Fourier.pdf

Fourier Analysis

Translation by Olof Staffans of the lecture notes

Fourier analyysi by Gustaf Gripenberg

January 5, 2009

0

Contents 0 Integration theory

3

1 Finite Fourier Transform

10

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

1.2 L2 -Theory (“Energy theory”) . . . . . . . . . . . . . . . . . . . .

14

1.3 Convolutions (”Faltningar”) . . . . . . . . . . . . . . . . . . . . .

21

1.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

1.4.1

Wirtinger’s Inequality . . . . . . . . . . . . . . . . . . . .

31

1.4.2

Weierstrass Approximation Theorem . . . . . . . . . . . .

32

1.4.3

Solution of Differential Equations . . . . . . . . . . . . . .

33

1.4.4

Solution of Partial Differential Equations . . . . . . . . . .

35

2 Fourier Integrals

36

1

2.1 L -Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

36

2.2 Rapidly Decaying Test Functions . . . . . . . . . . . . . . . . . .

43

2.3 L2 -Theory for Fourier Integrals . . . . . . . . . . . . . . . . . . .

45

2.4 An Inversion Theorem . . . . . . . . . . . . . . . . . . . . . . . .

48

2.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

52

2.5.1 2.5.2

The Poisson Summation Formula . . . . . . . . . . . . . . 52 1 (R) = C (R) ? . . . . . . . . . . . . . . . . . . . . . . Is Ld 53 0

2.5.3

The Euler-MacLauren Summation Formula . . . . . . . . .

54

2.5.4

Schwartz inequality . . . . . . . . . . . . . . . . . . . . . .

55

2.5.5

Heisenberg’s Uncertainty Principle . . . . . . . . . . . . .

55

2.5.6

Weierstrass’ Non-Differentiable Function . . . . . . . . . .

56

2.5.7

Differential Equations . . . . . . . . . . . . . . . . . . . .

59

2.5.8

Heat equation . . . . . . . . . . . . . . . . . . . . . . . . .

63

2.5.9

Wave equation . . . . . . . . . . . . . . . . . . . . . . . .

64

1

CONTENTS

3 Fourier Transforms of Distributions

2

67

3.1 What is a Measure? . . . . . . . . . . . . . . . . . . . . . . . . . .

67

3.2 What is a Distribution? . . . . . . . . . . . . . . . . . . . . . . .

69

3.3 How to Interpret a Function as a Distribution? . . . . . . . . . . .

71

3.4 Calculus with Distributions . . . . . . . . . . . . . . . . . . . . .

73

3.5 The Fourier Transform of a Distribution . . . . . . . . . . . . . . 76 3.6 The Fourier Transform of a Derivative . . . . . . . . . . . . . . .

77

3.7 Convolutions (”Faltningar”) . . . . . . . . . . . . . . . . . . . . .

79

3.8 Convergence in S ′ . . . . . . . . . . . . . . . . . . . . . . . . . . .

85

3.9 Distribution Solutions of ODE:s . . . . . . . . . . . . . . . . . . .

87

3.10 The Support and Spectrum of a Distribution . . . . . . . . . . . .

89

3.11 Trigonometric Polynomials . . . . . . . . . . . . . . . . . . . . . .

93

3.12 Singular differential equations . . . . . . . . . . . . . . . . . . . .

95

4 Fourier Series 5 The Discrete Fourier Transform

99 102

5.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.2 FFT=the Fast Fourier Transform . . . . . . . . . . . . . . . . . . 104 5.3 Computation of the Fourier Coefficients of a Periodic Function . . 107 5.4 Trigonometric Interpolation . . . . . . . . . . . . . . . . . . . . . 113 5.5 Generating Functions . . . . . . . . . . . . . . . . . . . . . . . . . 115 5.6 One-Sided Sequences . . . . . . . . . . . . . . . . . . . . . . . . . 116 5.7 The Polynomial Interpretation of a Finite Sequence . . . . . . . . 118 5.8 Formal Power Series and Analytic Functions . . . . . . . . . . . . 119 5.9 Inversion of (Formal) Power Series . . . . . . . . . . . . . . . . . . 121 5.10 Multidimensional FFT . . . . . . . . . . . . . . . . . . . . . . . . 122 6 The Laplace Transform

124

6.1 General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 6.2 The Standard Laplace Transform . . . . . . . . . . . . . . . . . . 125 6.3 The Connection with the Fourier Transform . . . . . . . . . . . . 126 6.4 The Laplace Transform of a Distribution . . . . . . . . . . . . . . 129 6.5 Discrete Time: Z-transform . . . . . . . . . . . . . . . . . . . . . 130 6.6 Using Laguerra Functions and FFT to Compute Laplace Transforms131

Chapter 0 Integration theory This is a short summary of Lebesgue integration theory, which will be used in the course. Fact 0.1. Some subsets (=“delm¨angder”) E ⊂ R = (−∞, ∞) are “measurable”

(=“m¨atbara”) in the Lebesgue sense, others are not.

General Assumption 0.2. All the subsets E which we shall encounter in this course are measurable. Fact 0.3. All measurable subsets E ⊂ R have a measure (=“m˚ att”) m(E), which

in simple cases correspond to “ the total length” of the set. E.g., the measure of

the interval (a, b) is b − a (and so is the measure of [a, b] and [a, b)). Fact 0.4. Some sets E have measure zero, i.e., m(E) = 0. True for example if E consists of finitely many (or countably many) points. (“m˚ attet noll”) The expression a.e. = “almost everywhere” (n.¨o. = n¨astan ¨overallt) means that something is true for all x ∈ R, except for those x which belong to some set E with measure zero. For example, the function ( 1, |x| ≤ 1 f (x) = 0, |x| > 1

is continuous almost everywhere. The expression fn (x) → f (x) a.e. means that the measure of the set x ∈ R for which fn (x) 9 f (x) is zero.

Think: “In all but finitely many points” (this is a simplification). 3

4

CHAPTER 0. INTEGRATION THEORY

Notation 0.5. R = (−∞, ∞), C = complex plane. The set of Riemann integrable functions f : I 7→ C (I ⊆ R is an interval) such that

Z

I

|f (x)|p dx < ∞, 1 ≤ p < ∞,

though much larger than the space C(I) of continuous functions on I, is not big enough for our purposes. This defect can be remedied by the use of the Lebesgue integral instead of the Riemann integral. The Lebesgue integral is more complicated to define and develop than the Riemann integral, but as a tool it is easier to use as it has better properties. The main difference between the Riemann and the Lebesgue integral is that the former uses intervals and their lengths while the latter uses more general point sets and their measures. Definition 0.6. A function f : I 7→ C (I ∈ R is an interval) is measurable if there exists a sequence of continuous functions fn so that fn (x) → f (x) for almost all x ∈ I (i.e., the set of points x ∈ I for which fn (x) 9 f (x) has measure zero). General Assumption 0.7. All the functions that we shall encounter in this course are measurable. Thus, the word “measurable” is understood throughout (when needed). Definition 0.8. Let 1 ≤ p < ∞, and I ⊂ R an interval. We write f ∈ Lp (I) if (f is measurable and)

Z

I

p

|f (x)|p dx < ∞.

We define the norm of f in L (I) to be kf kLp (I) =

Z

1/p |f (x)| dx . p

I

Physical interpretation: p=1

kf kL1 (I) =

Z

I

|f (x)|dx

5

CHAPTER 0. INTEGRATION THEORY

= “the total mass”. “Probability density” if f (x) ≥ 0, or a “size of the total population”.

p=2

kf kL2 (I) =

Z

1/2 |f (x)| dx 2

I

= “total energy” (e.g. in an electrical signal, such as alternating current). These two cases are the two important ones (we ignore the rest). The third important case is p = ∞. Definition 0.9. f ∈ L∞ (I) if (f is measurable and) there exists a number M < ∞ such that

|f (x)| < M

a.e.

The norm of f is kf kL∞ (I) = inf{M : |f (x)| ≤ M a.e.}, and it is denoted by kf kL∞ (I) = ess sup|f (x)| x∈I

(“essential supremum”, ”v¨asentligt supremum”). Think: kf kL∞ (I) = “the largest value of f in I if we ignore a set of measure zero”. For example:

f (x) =

⇒ kf kL∞ (I) = 1.

    0, x < 0

2, x = 0    1, x > 0

Definition 0.10. CC∞ (R) = D = the set of (real or complex-valued) functions on R which can be differentiated as many times as you wish, and which vanish outside of a bounded interval (such functions do exist!). CC∞ (I) = the same thing, but the function vanish outside of I. Infinitely many derivatives

=0

=0 I

6

CHAPTER 0. INTEGRATION THEORY

Theorem 0.11. Let I ⊂ R be an interval. Then CC∞ (I) is dense in Lp (I) for

all p, 1 ≤ p < ∞ (but not in L∞ (I)). That is, for every f ∈ Lp (I) it is possible to find a sequence fn ∈ CC∞ (I) so that

lim kfn − f kLp (I) = 0.

n→∞

Proof. “Straightforward” (but takes a lot of work).



Theorem 0.12 (Fatou’s lemma). Let fn (x) ≥ 0 and let fn (x) → f (x) a.e. as

n → ∞. Then

Z

I

f (x)dx ≤ lim

n→∞

Z

fn (x)dx

I

(if the latter limit exists). Thus, Z Z h i fn (x)dx lim fn (x) dx ≤ lim n→∞

I

n→∞

I

if fn ≥ 0 (“f can have no more total mass than fn , but it may have less”). Often

we have equality, but not always. Ex.

fn (x) =

(

n, 0 ≤ x ≤ 1/n 0, otherwise.

Homework: Compute the limits above in this case. Theorem 0.13 (Monotone Convergence Theorem). If 0 ≤ f1 (x) ≤ f2 (x) ≤ . . . and fn (x) → f (x) a.e., then Z Z f (x)dx = lim fn (x)dx (≤ ∞). n→∞

I

I

Thus, for a positive increasing sequence we have Z Z h i lim fn (x) dx = lim fn (x)dx n→∞

I

n→∞

I

(the mass of the limit is the limit of the masses).

Theorem 0.14 (Lebesgue’s dominated convergence theorem). (Extremely useful) If fn (x) → f (x) a.e. and |fn (x)| ≤ g(x) a.e. and Z g(x)dx < ∞ (i.e., g ∈ L1 (I)), I

7

CHAPTER 0. INTEGRATION THEORY

then

Z

Z h

f (x)dx = I

I

i

lim fn (x) dx = lim

n→∞

n→∞

Z

fn (x)dx.

I

Theorem 0.15 (Fubini’s theorem). (Very useful for multiple integrals). If f (is measurable and) Z Z I

then the double integral

J

|f (x, y)|dy dx < ∞

ZZ

f (x, y)dy dx

I×J

is well-defined, and equal to

Z

=

x∈I

Z

=

y∈J

Z

y∈J

Z

x∈I



f (x, y)dy dx  f (x, y)dx dy

If f ≥ 0, then all three integrals are well-defined, possibly = ∞, and if one of

them is < ∞, then so are the others, and they are equal.

Note: These theorems are very useful, and often easier to use than the corresponding theorems based on the Rieman integral. Theorem 0.16 (Integration by parts `a la Lebesgue). Let [a, b]be a finite interval, u ∈ L1 ([a, b]), v ∈ L1 ([a, b]), Z t Z t U(t) = U(a) + u(s)ds, V (t) = V (a) + v(s)ds, t ∈ [a, b]. a

Then

Z

a

b

u(t)V (t)dt = [U(t)V

a

Proof. Z

Z

b

u(t)V (t)

=

a

a

Fubini

=

Since Z b s

u(t)dt =

Z

a

b

−

Z s a

b

u(t)

Z

(t)]ba

−

Z

b

U(t)v(t)dt.

a

t

v(s)dsdt

a




U(b) − U(a) V (a) +

Z b Z a

u(t)dt = U(b) − U(a) −

Z

a

b s

 u(t)dt v(s)ds.

s

u(t)dt = U(b) − U(s),

8

CHAPTER 0. INTEGRATION THEORY

we get Z b u(t)V (t)dt = a




U(b) − U(a) V (a) +

Z

b

(U(b) − U(s)) v(s)ds Z b

= U(b) − U(a) V (a) + U(b) V (b) − V (a) − U(s)v(s)ds a Z b = U(b)V (b) − U(a)V (a) − U(s)v(s)ds.  a

a

Example 0.17. Sometimes we need test functions with special properties. Let us take a look how one can proceed. b(t)

1

b(t) =

(

t

1

e− t(1−t) , 0 < t < 1 0

, otherwise.

Then we can show that b ∈ C ∞ (R), and b is a test function with compact support. Rt B(t) . Let B(t) = −∞ b(s)ds and norm it F (t) = B(1) F(t)

1

1

F (t) =

   

0

t

,t ≤ 0

1 ,t ≥ 1    increase , 0 < t < 1.

Further F (t) + F (t − 1) = 1, ∀t ∈ R, clearly true for t ≤ 0 and t ≥ 1. For 0 < t < 1 we check the derivative

d 1 1 (F (t) − F (1 − t)) = [B ′ (t) − B ′ (1 − t)] = [b(t) − b(1 − t)] = 0. dt B(1) B(1)

9

CHAPTER 0. INTEGRATION THEORY

F(1−t)

F(t)

Let G(t) = F (Nt). Then G increases from 0 to 1 on the interval 0 ≤ t ≤ G(t) + G(

G(1/N−t)

1 − t) = F (Nt) − F (1 − Nt) = 1, ∀t ∈ R. N

G(t)

1/N

1

1 . N

Chapter 1 The Fourier Series of a Periodic Function 1.1

Introduction

Notation 1.1. We use the letter T with a double meaning: a) T = [0, 1) b) In the notations Lp (T), C(T), C n (T) and C ∞ (T) we use the letter T to imply that the functions are periodic with period 1, i.e., f (t + 1) = f (t) for all t ∈ R. In particular, in the continuous case we require f (1) = f (0).

Since the functions are periodic we know the whole function as soon as we know the values for t ∈ [0, 1). R 1/p 1 Notation 1.2. kf kLp (T) = 0 |f (t)|p dt , 1 ≤ p < ∞. kf kC(T) = maxt∈T |f (t)|

(f continuous).

Definition 1.3. f ∈ L1 (T) has the Fourier coefficients Z 1 ˆ f (n) = e−2πint f (t)dt, n ∈ Z, 0

ˆ where Z = {0, ±1, ±2, . . . }. The sequence {f(n)} n∈Z is the (finite) Fourier

transform of f . Note:

fˆ(n) =

Z

s+1 s

e−2πint f (t)dt ∀s ∈ R, 10

11

CHAPTER 1. FINITE FOURIER TRANSFORM

since the function inside the integral is periodic with period 1. Note: The Fourier transform of a periodic function is a discrete sequence. Theorem 1.4. i) |fˆ(n)| ≤ kf kL1 (T) ,

∀n ∈ Z

ii) limn→±∞ fˆ(n) = 0. Note: ii) is called the Riemann–Lebesgue lemma. Proof. i) |fˆ(n)| = |

R1 0

e−2πint f (t)dt| ≤

R1 0

|e−2πint f (t)|dt =

the triangle inequality for integrals).

R1 0

|f (t)|dt = kf kL1 (T) (by

ii) First consider the case where f is continuously differentiable, with f (0) = f (1). Then integration by parts gives Z 1 ˆ f(n) = e−2πint f (t)dt 0

Z 1 1 1  −2πint 1 = e f (t) 0 + e−2πint f ′ (t)dt −2πin 2πin 0 1 ˆ′ f (n), so by i), = 0+ 2πin Z 1 1 1 ˆ |f(n)| = |fˆ′(n)| ≤ |f ′(s)|ds → 0 as n → ∞. 2πn 2πn 0 In the general case, take f ∈ L1 (T) and ε > 0. By Theorem 0.11 we can choose some g which is continuously differentiable with g(0) = g(1) = 0 so that kf − gkL1 (T) = By i),

Z

1 0

|f (t) − g(t)|dt ≤ ε/2.

ˆ ˆ − gˆ(n) + gˆ(n)| |f(n)| = |f(n)

ˆ − gˆ(n)| + |ˆ ≤ |f(n) g (n)| ≤ kf − gkL1 (T) + |ˆ g (n)| ≤ ǫ/2 + |ˆ g (n)|.

By the first part of the proof, for n large enough, |ˆ g (n)| ≤ ε/2, and so ˆ |f(n)| ≤ ε. ˆ This shows that |f(n)| → 0 as n → ∞.



12

CHAPTER 1. FINITE FOURIER TRANSFORM

Question 1.5. If we know {fˆ(n)}∞ n=−∞ then can we reconstruct f (t)? Answer: is more or less ”Yes”.

Definition 1.6. C n (T) = n times continuously differentiable functions, periodic with period 1. (In particular, f (k) (1) = f (k) (0) for 0 ≤ k ≤ n.) Theorem 1.7. For all f ∈ C 1 (T) we have f (t) = lim

N X

N →∞ M →∞ n=−M

fˆ(n)e2πint ,

t ∈ R.

(1.1)

We shall see later that the convergence is actually uniform in t. Proof. Step 1. We shift the argument of f by replacing f (s) by g(s) = f (s+t). Then ˆ gˆ(n) = e2πint f(n), and (1.1) becomes f (t) = g(0) = lim

N X

N →∞ M →∞ n=−M

fˆ(n)e2πint .

Thus, it suffices to prove the case where t = 0 . Step 2: If g(s) is the constant function g(s) ≡ g(0) = f (t), then (1.1) holds since gˆ(0) = g(0) and gˆ(n) = 0 for n 6= 0 in this case. Replace g(s) by h(s) = g(s) − g(0). Then h satisfies all the assumptions which g does, and in addition, h(0) = 0. Thus it suffices to prove the case where both t = 0 and f (0) = 0. For simplicity we write f instead of h, but we suppose below that t = 0 and f (0) = 0 Step 2: Define g(s) =

(

f (s) , e−2πis −1 if ′ (0) , 2π

s 6= integer (=“heltal”)

s = integer.

For s = n = integer we have e−2πis − 1 = 0, and by l’Hospital’s rule f ′ (s) f ′ (s) = = g(n) s→0 −2πie−2πis −2πi

lim g(s) = lim

s→n

13

CHAPTER 1. FINITE FOURIER TRANSFORM

(since e−i2πn = 1). Thus g is continuous. We clearly have
f (s) = e−2πis − 1 g(s),

so fˆ(n) =

Z

e−2πins f (s)ds

(1.2)

(use (1.2))

ZT

e−2πins (e−2πis − 1)g(s)ds ZT Z −2πi(n+1)s e g(s)ds − e−2πins g(s)ds = =

T

T

= gˆ(n + 1) − gˆ(n).

Thus, N X

n=−M

fˆ(n) = gˆ(N + 1) − gˆ(−M) → 0

by the Rieman–Lebesgue lemma (Theorem 1.4)  By working a little bit harder we get the following stronger version of Theorem 1.7: Theorem 1.8. Let f ∈ L1 (T), t0 ∈ R, and suppose that Z t0 +1 f (t) − f (t0 ) dt < ∞. t − t0 t0 −1

Then

f (t0 ) = lim

N X

N →∞ M →∞ n=−M

fˆ(n)e2πint0

(1.3)

t∈R

Proof. We can repeat Steps 1 and 2 of the preceding proof to reduce the Theorem to the case where t0 = 0 and f (t0 ) = 0. In Step 3 we define the function g in the same way as before for s 6= n, but leave g(s) undefined for s = n.

Since lims→0 s−1 (e−2πis − 1) = −2πi 6= 0, the function g belongs to L1 (T) if and

only if condition (1.3) holds. The continuity of g was used only to ensure that

g ∈ L1 (T), and since g ∈ L1 (T) already under the weaker assumption (1.3), the rest of the proof remains valid without any further changes.



14

CHAPTER 1. FINITE FOURIER TRANSFORM

Summary 1.9. If f ∈ L1 (T), then the Fourier transform {fˆ(n)}∞ n=−∞ of f is well-defined, and fˆ(n) → 0 as n → ∞. If f ∈ C 1 (T), then we can reconstruct f from its Fourier transform through f (t) =

∞ X

fˆ(n)e2πint = lim

N X

N →∞ M →∞ n=−M

n=−∞

fˆ(n)e2πint

!

.

The same reconstruction formula remains valid under the weaker assumption of Theorem 1.8.

1.2

L2-Theory (“Energy theory”)

This theory is based on the fact that we can define an inner product (scalar product) in L2 (T), namely hf, gi =

Z

1

f (t)g(t)dt,

0

f, g ∈ L2 (T).

Scalar product means that for all f, g, h ∈ L2 (T) i) hf + g, hi = hf, hi + hg, hi ii) hλf, gi = λhf, gi ∀λ ∈ C iii) hg, f i = hf, gi (complex conjugation) iv) hf, f i ≥ 0, and = 0 only when f ≡ 0. These are the same rules that we know from the scalar products in Cn . In addition we have kf k2L2 (T)

=

Z

2

T

|f (t)| dt =

Z

T

f (t)f (t)dt = hf, f i.

This result can also be used to define the Fourier transform of a function f ∈ L2 (T), since L2 (T) ⊂ L1 (T).

Lemma 1.10. Every function f ∈ L2 (T) also belongs to L1 (T), and kf kL1 (T) ≤ kf kL2 (T) .

15

CHAPTER 1. FINITE FOURIER TRANSFORM

Proof. Interpret

R

T

|f (t)|dt as the inner product of |f (t)| and g(t) ≡ 1. By

Schwartz inequality (see course on Analysis II), Z Z |hf, gi| = |f (t)| · 1dt ≤ kf kL2 · kgkL2 = kf kL2 (T) 12 dt = kf kL2 (T) . T

T

Thus, kf kL1 (T) ≤ kf kL2 (T) . Therefore: Z 2 f ∈ L (t) =⇒ |f (t)|dt < ∞ T Z ˆ =⇒ f (n) = e−2πint f (t)dt is defined for all n. T

It is not true that L2 (R) ⊂ L1 (R). Counter example:   ∈ L2 (R)   1 f (t) = √ ∈ / L1 (R) 1 + t2    ∈ C ∞ (R)

(too large at ∞).

Notation 1.11. en (t) = e2πint , n ∈ Z, t ∈ R. Theorem 1.12 (Plancherel’s Theorem). Let f ∈ L2 (T). Then i)

P∞

ˆ

n=−∞ |f(n)|

ii) f =

P∞

n=−∞

2

=

R1 0

|f (t)|2 dt = kf k2L2 (T) ,

fˆ(n)en in L2 (T) (see explanation below).

Note: This is a very central result in, e.g., signal processing. P 2 2 ˆ Note: It follows from i) that the sum ∞ n=−∞ |f(n)| always converges if f ∈ L (T) Note: i) says that ∞ X

2 ˆ |f(n)| = the square of the total energy of the Fourier coefficients

n=−∞

= Z the square of the total energy of the original signal f = |f (t)|2 dt T

Note: Interpretation of ii): Define fM,N =

N X

n=−M

fˆ(n)en =

N X

n=−M

2πint ˆ f(n)e .

16

CHAPTER 1. FINITE FOURIER TRANSFORM

Then lim kf − fM,N k2 = 0 ⇐⇒

M →∞ N →∞

lim

M →∞ N →∞

Z

1

0

|f (t) − fM,N (t)|2 dt = 0

(fM,N (t) need not converge to f (t) at every point, and not even almost everywhere). The proof of Theorem 1.12 is based on some auxiliary results: P P∞ 2 Theorem 1.13. If gn ∈ L2 (T), fN = N n=0 gn , gn ⊥ gm , and n=0 kgn kL2 (T) < ∞, then the limit

f = lim

N →∞

exists in L2 .

N X

gn

n=0

Proof. Course on “Analysis II” and course on “Hilbert Spaces”.



Interpretation: Every orthogonal sum with finite total energy converges. P Lemma 1.14. Suppose that ∞ n=−∞ |c(n)| < ∞. Then the series ∞ X

c(n)e2πint

n=−∞

converges uniformly to a continuous limit function g(t). Proof. i) The series the limit

P∞

n=−∞

c(n)e2πint converges absolutely (since |e2πint | = 1), so g(t) =

∞ X

c(n)e2πint

n=−∞

exist for all t ∈ R.

ii) The convergens is uniform, because the error |

m X

n=−m

c(n)e2πint − g(t)| = | ≤ =

X

|n|>m

X

c(n)e2πint |

|c(n)e2πint |

|n|>m

X

|c(n)| → 0 as m → ∞.

|n|>m

17

CHAPTER 1. FINITE FOURIER TRANSFORM

iii) If a sequence of continuous functions converge uniformly, then the limit is continuous (proof “Analysis II”). proof of Theorem 1.12. (Outline) 0 ≤ kf − fM,N k2 = hf − fM,N , f − fM,N i = hf, f i − hfM,N , f i − hf, fM,N i + hfM,N , fM,N i | {z } | {z } | {z } | {z } I

II

III

IV

I = hf, f i = kf k2L2 (T) .

II = h =

N X

n=−M N X

n=−M

=

N X

n=−M

fˆ(n)en , f i =

ˆ f(n)hf, en i =

N X

n=−M N X

fˆ(n)hen , f i fˆ(n)fˆ(n)

n=−M

2 ˆ |f(n)| .

III = (the complex conjugate of II) = II. N N E D X X ˆ ˆ f (m)em f(n)en , IV = m=−M

n=−M

=

N X

n=−M

=

N X

n=−M

ˆ fˆ(m) hen , em i f(n) | {z } m δn

2 ˆ |f(n)| = II = III.

Thus, adding I − II − III + IV = I − II ≥ 0, i.e., kf k2L2 (T)

−

N X

n=−M

2 ˆ |f(n)| ≥ 0.

This proves Bessel’s inequality ∞ X

n=−∞

2 ˆ |f(n)| ≤ kf k2L2 (T) .

How do we get equality? By Theorem 1.13, applied to the sums N X n=0

fˆ(n)en and

−1 X

n=−M

fˆ(n)en ,

(1.4)

18

CHAPTER 1. FINITE FOURIER TRANSFORM

the limit

N X

g = lim fM,N = lim

M →∞ N →∞ n=−M

M →∞ N →∞

fˆ(n)en

(1.5)

does exist. Why is f = g? (This means that the sequence en is complete!). This is (in principle) done in the following way i) Argue as in the proof of Theorem 1.4 to show that if f ∈ C 2 (T), then |fˆ(n)| ≤ 1/(2πn)2 kf ′′ kL1 for n 6= 0. In particular, this means that P∞ ˆ n=−∞ |f(n)| < ∞. By Lemma 1.14, the convergence in (1.5) is actually uniform, and by Theorem 1.7, the limit is equal to f . Uniform convergence

implies convergence in L2 (T), so even if we interpret (1.5) in the L2 -sense, the limit is still equal to f a.e. This proves that fM,N → f in L2 (T) if f ∈ C 2 (T).

ii) Approximate an arbitrary f ∈ L2 (T) by a function h ∈ C 2 (T) so that kf − hkL2 (T) ≤ ε.

iii) Use i) and ii) to show that kf − gkL2 (T) ≤ ε, where g is the limit in (1.5). Since ε is arbitrary, we must have g = f . 

Definition 1.15. Let 1 ≤ p < ∞. p

ℓ (Z) = set of all sequences

{an }∞ n=−∞

satisfying

The norm of a sequence a ∈ ℓp (Z) is kakℓp (Z) =

∞ X

|an |p

n=−∞

∞ X

n=−∞

|an |p < ∞.

!1/p

Analogous to Lp (I): p = 1 kakℓ1 (Z) = ”total mass” (probability), p = 2 kakℓ2 (Z) = ”total energy”.

In the case of p = 2 we also define an inner product ha, bi =

∞ X

n=−∞

an bn .

CHAPTER 1. FINITE FOURIER TRANSFORM

19

Definition 1.16. ℓ∞ (Z) = set of all bounded sequences {an }∞ n=−∞ . The norm in ℓ∞ (Z) is

kakℓ∞ (Z) = sup|an |. n∈Z

For details: See course in ”Analysis II”. Definition 1.17. c0 (Z) = the set of all sequences {an }∞ n=−∞ satisfying limn→±∞ an = 0. We use the norm

kakc0 (Z) = max|an | n∈Z

in c0 (Z). Note that c0 (Z) ⊂ ℓ∞ (Z), and that kakc0 (Z) = kakℓ∞ (Z) if {a}∞ n=−∞ ∈ c0 (Z). Theorem 1.18. The Fourier transform maps L2 (T) one to one onto ℓ2 (Z), and the Fourier inversion formula (see Theorem 1.12 ii) maps ℓ2 (Z) one to one onto L2 (T). These two transforms preserves all distances and scalar products. Proof. (Outline) i) If f ∈ L2 (T) then fˆ ∈ ℓ2 (Z). This follows from Theorem 1.12. 2 ii) If {an }∞ n=−∞ ∈ ℓ (Z), then the series N X

an e2πint

n=−M

converges to some limit function f ∈ L2 (T). This follows from Theorem

1.13.

ˆ iii) If we compute the Fourier coefficients of f , then we find that an = f(n). Thus, {an }∞ n=−∞ is the Fourier transform of f . This shows that the Fourier transform maps L2 (T) onto ℓ2 (Z).

iv) Distances are preserved. If f ∈ L2 (T), g ∈ L2 (T), then by Theorem 1.12 i),

kf − gkL2 (T) = kfˆ(n) − gˆ(n)kℓ2 (Z) , i.e.,

Z

T

2

|f (t) − g(t)| dt =

∞ X

|fˆ(n) − gˆ(n)|2 .

n=−∞

CHAPTER 1. FINITE FOURIER TRANSFORM

20

v) Inner products are preserved: Z |f (t) − g(t)|2 dt = hf − g, f − gi T

= hf, f i − hf, gi − hg, f i + hg, gi = hf, f i − hf, gi − hf, gi + hg, gi = hf, f i + hg, gi − 2ℜRehf, gi.

In the same way, ∞ X

|fˆ(n) − gˆ(n)|2 = hfˆ − gˆ, fˆ − gˆi

n=−∞

ˆ + hˆ = hfˆ, fi g , gˆi − 2ℜhfˆ, gˆi.

By iv), subtracting these two equations from each other we get ℜhf, gi = ℜhfˆ, gˆi. If we replace f by if , then Imhf, gi = Re ihf, gi = ℜhif, gi

ˆ gˆi = Re ihfˆ, gˆi = ℜhif,

= Imhfˆ, gˆi.

Thus, hf, giL2 (R) = hfˆ, gˆiℓ2 (Z) , or more explicitely, Z

f (t)g(t)dt =

T

∞ X

g (n). fˆ(n)ˆ

(1.6)

n=−∞

This is called Parseval’s identity. Theorem 1.19. The Fourier transform maps L1 (T) into c0 (Z) (but not onto), and it is a contraction, i.e., the norm of the image is ≤ the norm of the original

function.

Proof. This is a rewritten version of Theorem 1.4. Parts i) and ii) say that ˆ {fˆ(n)}∞  n=−∞ ∈ c0 (Z), and part i) says that kf (n)kc0 (Z) ≤ kf kL1 (T) . The proof that there exist sequences in c0 (Z) which are not the Fourier transform of some function f ∈ L1 (T) is much more complicated.

CHAPTER 1. FINITE FOURIER TRANSFORM

1.3

21

Convolutions (”Faltningar”)

Definition 1.20. The convolution (”faltningen”) of two functions f, g ∈ L1 (T) is

where

R

T

=

R α+1 α

(f ∗ g)(t) =

Z

T

f (t − s)g(s)ds,

for all α ∈ R, since the function s 7→ f (t − s)g(s) is periodic.

Note: In this integral we need values of f and g outside of the interval [0, 1), and therefore the periodicity of f and g is important. Theorem 1.21. If f, g ∈ L1 (T), then (f ∗ g)(t) is defined almost everywhere,

and f ∗ g ∈ L1 (T). Furthermore,





f ∗ g 1 ≤ f 1 g 1 L (T) L (T) L (T)

Proof. (We ignore measurability) We begin with (1.7)

f ∗ g 1 L (T)

Z

=

|(f ∗ g)(t)| dt Z Z f (t − s)g(s) ds dt ZT TZ |f (t − s)g(s)| ds dt t∈T s∈T  Z Z |f (t − s)| dt |g(s)| ds s∈T t∈T Z  Z |f (v)| dv |g(s)| ds s∈T v∈T | {z } T

= △-ineq.

≤

Fubini

=

Put v=t−s,dv=dt

=

=kf kL1 (T)

=

kf kL1 (T)

Z

s∈T

|g(s)| ds = kf kL1 (T) kgkL1 (T)

This integral is finite. By Fubini’s Theorem 0.15 Z f (t − s)g(s)ds T

is defined for almost all t.



Theorem 1.22. For all f, g ∈ L1 (T) we have ˆ g (n), n ∈ Z (f[ ∗ g)(n) = f(n)ˆ

(1.7)

22

CHAPTER 1. FINITE FOURIER TRANSFORM

Proof. Homework. Thus, the Fourier transform maps convolution onto pointwise multiplication. Theorem 1.23. If k ∈ C n (T) (n times continuously differentiable) and f ∈ L1 (T), then k ∗ f ∈ C n (T), and (k ∗ f )(m) (t) = (k (m) ∗ f )(t) for all m = 0, 1, 2, . . . , n.

Proof. (Outline) We have for all h > 0 Z 1 1 1 [k(t + h − s) − k(t − s)] f (s)ds. [(k ∗ f ) (t + h) − (k ∗ f )(t)] = h h 0 By the mean value theorem, k(t + h − s) = k(t − s) + hk ′ (ξ), for some ξ ∈ [t − s, t − s + h], and h1 [k(t + h − s) − k(t − s)] = f (ξ) → k ′ (t − s) as h → 0, and 1 [k(t + h − s) − k(t − s)] = |f ′ (ξ)| ≤ M, where M = supT |k ′ (s)|. By h

the Lebesgue dominated convergence theorem (which is true also if we replace n → ∞ by h → 0)(take g(x) = M|f (x)|) Z 1 Z 1 1 lim [k(t + h − s) − k(t − s)]f (s) ds = k ′ (t − s)f (s) ds, h→0 0 h 0

so k ∗ f is differentiable, and (k ∗ f )′ = k ′ ∗ f . By repeating this n times we find

that k ∗ f is n times differentiable, and that (k ∗ f )(n) = k (n) ∗ f . We must still

show that k (n) ∗ f is continuous. This follows from the next lemma.



Lemma 1.24. If k ∈ C(T) and f ∈ L1 (T), then k ∗ f ∈ C(T). Proof. By Lebesgue dominated convergence theorem (take g(t) = 2kkkC(T) f (t)), (k ∗ f )(t + h) − (k ∗ f )(t) =

Z

0

1

[k(t + h − s) − k(t − s)]f (s)ds → 0 as h → 0.

Corollary 1.25. If k ∈ C 1 (T) and f ∈ L1 (T), then for all t ∈ R (k ∗ f )(t) =

∞ X

ˆ fˆ(n). e2πint k(n)

n=−∞

Proof. Combine Theorems 1.7, 1.22 and 1.23. ˆ so Interpretation: This is a generalised inversion formula. If we choose k(n) that

23

CHAPTER 1. FINITE FOURIER TRANSFORM

ˆ i) k(n) ≈ 1 for small |n| ˆ ii) k(n) ≈ 0 for large |n|, then we set a “filtered” approximation of f , where the “high frequences” (= high values of |n|) have been damped but the “low frequences” (= low values of |n|) ˆ remain. If we can take k(n) = 1 for all n then this gives us back f itself, but this is impossible because of the Riemann-Lebesgue lemma. Problem: Find a “good” function k ∈ C 1 (T) of this type. ˆ Solution: “The Fejer kernel” is one possibility. Choose k(n) to be a “triangular function”: F4 (n)

n=0

n=4

Fix m = 0, 1, 2 . . . , and define Fˆm (n) =

(

m+1−|n| m+1

0

, |n| ≤ m

, |n| > m

(6= 0 in 2m + 1 points.) We get the corresponding time domain function Fm (t) by using the invertion formula: Fm (t) =

m X

Fˆm (n)e2πint .

n=−m

Theorem 1.26. The function Fm (t) is explicitly given by 1 sin2 ((m + 1)πt) Fm (t) = . m+1 sin2 (πt) Proof. We are going to show that j m X X

j=0 n=−j

e2πint =

sin(π(m + 1)t)
2 sin πt

when t 6= 0.

24

CHAPTER 1. FINITE FOURIER TRANSFORM

Let z = e2πit , z = e−2πit , for t 6= n, n = 0, 1, 2, . . . . Also z 6= 1, and j X

2πint

e

=

j X

2πint

e

+

j+1

=

1−z 1−z j

=

−2πint

e

=

j X

z +

z −z . 1−z

zn

=1

1−z z(1 − z ) = 1−z 1−z

j+1

j X n=1

j+1

j

+

n

n=0

n=1

n=0

n=−j

j X

z}|{ z · z(1 − z j ) + z − |{z} z·z =1

Hence j m X X

e2πint

j=0 n=−j

m m X X 1 = = zj − z j+1 1 − z 1 − z j=0 j=0 j=0    m+1 m+1 1 1−z 1−z = −z 1−z 1−z 1−z m X z j − z j+1

!

=1

z}|{ z · z(1 − z m+1 i 1 h 1 − z m+1 − = 1−z 1−z z(1 − z) h m+1 1−z 1 − z m+1 i 1 − = 1−z 1−z z−1 m+1 m+1 −z +2−z = . |1 − z|2 sin t =

1 (eit 2i

− e−it ), cos t = 21 (eit + e−it ). Now

|1 − z| = |1 − e2πit | = |eiπt (e−iπt − eiπt )| = |e−iπt − eiπt | = 2|sin(πt)| and

Hence

z m+1 − 2 + z m+1 = e2πi(m+1) − 2 + e−2πi(m+1)
2
2 = eπi(m+1) − e−πi(m+1) = 2i sin(π(m + 1)) . j m X X

e2πint =

j=0 n=−j

4(sin(π(m + 1)))2  sin(π(m + 1)) 2 = 4(sin(πt))2 sin(πt)

Note also that

j m X X

j=0 n=−j

2πint

e

=

m m X X

n=−m j=|n|

2πint

e

=

m X

(m + 1 − |n|)e2πint .

n=−m

25

CHAPTER 1. FINITE FOURIER TRANSFORM

6 5 4 3 2 1

-2

-1

1

2

Comment 1.27. i) Fm (t) ∈ C ∞ (T) (infinitely many derivatives). ii) Fm (t) ≥ 0. iii)

R

|F (t)|dt = T m

R

T

Fm (t)dt = Fˆm (0) = 1,

so the total mass of Fm is 1.

iv) For all δ, 0 < δ < 21 , lim

m→∞

Z

1−δ

Fm (t)dt = 0,

δ

i.e. the mass of Fm gets concentrated to the integers t = 0, ±1, ±2 . . .

as m → ∞.

Definition 1.28. A sequence of functions Fm with the properties i)-iv) above is called a (periodic) approximate identity. (Often i) is replaced by Fm ∈ L1 (T).) Theorem 1.29. If f ∈ L1 (T), then, as m → ∞, i) Fm ∗ f → f in L1 (T), and ii) (Fm ∗ f )(t) → f (t) for almost all t. Here i) means that

R

T

|(Fm ∗ f )(t) − f (t)|dt → 0 as m → ∞

Proof. See page 27. By combining Theorem 1.23 and Comment 1.27 we find that Fm ∗ f ∈ C ∞ (T).

This combined with Theorem 1.29 gives us the following periodic version of Theorem 0.11:

26

CHAPTER 1. FINITE FOURIER TRANSFORM

Corollary 1.30. For every f ∈ L1 (T) and ε > 0 there is a function g ∈ C ∞ (T) such that kg − f kL1 (T) ≤ ε.

Proof. Choose g = Fm ∗ f where m is large enough.



To prove Theorem 1.29 we need a number of simpler results: Lemma 1.31. For all f, g ∈ L1 (T) we have f ∗ g = g ∗ f Proof. (f ∗ g)(t)

= t−s=v,ds=−dv

=

Z

ZT

T

f (t − s)g(s)ds f (v)g(t − v)dv = (g ∗ f )(t)



We also need: Theorem 1.32. If g ∈ C(T), then Fm ∗ g → g uniformly as m → ∞, i.e. max|(Fm ∗ g)(t) − g(t)| → 0 as m → ∞. t∈R

Proof. (Fm ∗ g)(t) − g(t)

Lemma 1.31

=

(g ∗ Fm )(t) − g(t)

Comment 1.27

=

Z

(g ∗ Fm )(t) − g(t) Fm (s)ds T Z [g(t − s) − g(t)]Fm (s)ds.

=

T

Since g is continuous and periodic, it is uniformly continuous, and given ε > 0 there is a δ > 0 so that |g(t − s) − g(t)| ≤ ε if |s| ≤ δ. Split the intgral above

into (choose the interval of integration to be [− 21 , 21 ]) Z

1 2

− 21

[g(t − s) − g(t)]Fm (s)ds =

Z

−δ

+

−1

Z

δ

+

−δ |{z}

2 |{z}

II

I

Z

1 2

δ |{z} III

 [g(t − s) − g(t)]Fm (s)ds

Let M = supt∈R |g(t)|. Then |g(t − s) − g(t)| ≤ 2M, and Z Z 1! |I + III| ≤

−δ

− 21

= 2M

Z

δ

2

+

2MFm (s)ds

δ

1−δ

Fm (s)ds

27

CHAPTER 1. FINITE FOURIER TRANSFORM

and by Comment 1.27iv) this goes to zero as m → ∞. Therefore, we can choose m so large that

|I + III| ≤ ε (m ≥ m0 , and m0 large.) |II| ≤

Z

≤ ε ≤ ε

δ −δ

Z

|g(t − s) − g(t)|Fm (s)ds δ

Fm (s)ds

−δ

Z

1 2

− 12

Fm (s)ds = ε

Thus, for m ≥ m0 we have |(Fm ∗ g)(t) − g(t)| ≤ 2ε (for all t). Thus, limm→∞ supt∈R |(Fm ∗ g)(t) − g(t)| = 0, i.e., (Fm ∗ g)(t) → g(t) uniformly as m → ∞.



The proof of Theorem 1.29 also uses the following weaker version of Lemma 0.11: Lemma 1.33. For every f ∈ L1 (T) and ε > 0 there is a function g ∈ C(T) such

that kf − gkL1 (T) ≤ ε.

Proof. Course in Lebesgue integration theory.



(We already used a stronger version of this lemma in the proof of Theorem 1.12.) Proof of Theorem 1.29, part i): (The proof of part ii) is bypassed, typically proved in a course on integration theory.) Let ε > 0, and choose some g ∈ C(T) with kf − gkL1 (T) ≤ ε. Then kFm ∗ f − f kL1 (T)

≤

kFm ∗ g − g + Fm ∗ (f − g) − (f − g)kL1(t)

≤

kFm ∗ g − gkL1(T) + kFm ∗ (f − g)kL1(T) + k(f − g)kL1 (T)

≤

kFm ∗ g − gkL1(T) + (kFm kL1 (T) + 1) kf − gkL1(T) {z } | {z } |

Thm 1.21

=2

=

kFm ∗ g − gkL1(T) + 2ε.

≤ε

28

CHAPTER 1. FINITE FOURIER TRANSFORM

Now kFm ∗ g − gkL1 (T) = ≤

Z

1

|(Fm ∗ g(t) − g(t)| dt

0

Z

1

max |(Fm ∗ g(s) − g(s)| dt Z 1 = max |(Fm ∗ g(s) − g(s)| · dt . s∈[0,1] 0 | {z } 0 s∈[0,1]

=1

By Theorem 1.32, this tends to zero as m → ∞. Thus for large enough m, kFm ∗ f − f kL1 (T) ≤ 3ε, so Fm ∗ f → f in L1 (T) as m → ∞.



(Thus, we have “almost” proved Theorem 1.29 i): we have reduced it to a proof of Lemma 1.33 and other “standard properties of integrals”.) In the proof of Theorem 1.29 we used the “trivial” triangle inequality in L1 (T): Z Z kf + gkL1(T) = |f (t) + g(t)| dt ≤ |f (t)| + |g(t)| dt = kf kL1 (T) + kgkL1(T)

Similar inequalities are true in all Lp (T), 1 ≤ p ≤ ∞, and a more “sophisticated” version of the preceding proof gives:

Theorem 1.34. If 1 ≤ p < ∞ and f ∈ Lp (T), then Fm ∗ f → f in Lp (T) as m → ∞, and also pointwise a.e. Proof. See Gripenberg. Note: This is not true in L∞ (T). The correct “L∞ -version” is given in Theorem 1.32. Corollary 1.35. (Important!) If f ∈ Lp (T), 1 ≤ p < ∞, or f ∈ C n (T), then m X m + 1 − |n| ˆ lim f (n)e2πint = f (t), m→∞ m+1 n=−m

where the convergence is in the norm of Lp , and also pointwise a.e. In the case where f ∈ C n (T) we have uniform convergence, and the derivatives of order ≤ n

also converge uniformly.

29

CHAPTER 1. FINITE FOURIER TRANSFORM

Proof. By Corollary 1.25 and Comment 1.27, m X m + 1 − |n| ˆ f (n)e2πint = (Fm ∗ f )(t) m+1 n=−m

The rest follows from Theorems 1.34, 1.32, and 1.23, and Lemma 1.31.  Interpretation: We improve the convergence of the sum ∞ X

2πint ˆ f(n)e

n=−∞

by multiplying the coefficients by the “damping factors”

m+1−|n| , |n| m+1

≤ m. This

particular method is called C´esaro summability. (Other “summability” methods use other damping factors.) Theorem 1.36. (Important!) The Fourier coefficients fˆ(n), n ∈ Z of a function f ∈ L1 (T) determine f uniquely a.e., i.e., if fˆ(n) = gˆ(n) for all n, then f (t) = g(t) a.e.

ˆ Proof. Suppose that gˆ(n) = f(n) for all n. Define h(t) = f (t) − g(t). Then ˆ h(n) = fˆ(n) − gˆ(n) = 0, n ∈ Z. By Theorem 1.29, m X m + 1 − |n| ˆ h(t) = lim h(n) e2πint = 0 m→∞ m + 1 |{z} n=−m =0

in the “L1 -sense”, i.e.

khk =

Z

1

0

|h(t)| dt = 0

This implies h(t) = 0 a.e., so f (t) = g(t) a.e.



Theorem 1.37. Suppose that f ∈ L1 (T) and that

series

∞ X

P∞

ˆ

n=−∞ |f(n)|

< ∞. Then the

2πint ˆ f(n)e

n=−∞

converges uniformly to a continuous limit function g(t), and f (t) = g(t) a.e. Proof. The uniform convergence follows from Lemma 1.14. We must have f (t) = g(t) a.e. because of Theorems 1.29 and 1.36. The following theorem is much more surprising. It says that not every sequence {an }n∈Z is the set of Fourier coefficients of some f ∈ L1 (T).

30

CHAPTER 1. FINITE FOURIER TRANSFORM

Theorem 1.38. Let f ∈ L1 (T), fˆ(n) ≥ 0 for n ≥ 0, and fˆ(−n) = −fˆ(n) (i.e. fˆ(n) is an odd function). Then ∞ X 1ˆ i) f (n) < ∞ n n=1 ii)

∞ X 1ˆ | f(n)| < ∞. n n=−∞ n6=0

Proof. Second half easy: Since fˆ is odd, X 1 X 1 X 1 | fˆ(n)| + | fˆ(−n)| | fˆ(n)| = n n n n>0 n<0 n6=0 n∈Z

= 2

∞ X 1 | fˆ(n)| < ∞ n n=1

if i) holds.

Rt ˆ ˆ i): Note that f(n) = −fˆ(−n) gives f(0) = 0. Define g(t) = 0 f (s)ds. Then R1 g(1) − g(0) = 0 f (s)ds = fˆ(0) = 0, so that g is continuous. It is not difficult to show (=homework) that

1 ˆ f (n), n 6= 0. 2πin

gˆ(n) = By Corollary 1.35,

m X m + 1 − |n| gˆ(n) e|2πin0 m→∞ |{z} {z } m + 1 {z } n=−m |

g(0) = gˆ(0) |e2πi·0·0 {z } + lim =1

even

even

=1

m X 2 m + 1 − n fˆ(n) . lim 2πi m→∞ n=0 m + 1 | {z n}

= gˆ(0) +

≥0

Thus lim

m→∞

m X m + 1 − n fˆ(n) n=1

m+1

n

= K = a finite pos. number.

In particular, for all finite M, M ˆ X f (n) n=1

and so

P∞

fˆ(n) n=1 n

n

= lim

≤ K < ∞.

m→∞

M X m + 1 − n fˆ(n) n=1

m+1

n

≤ K,



ˆ Theorem 1.39. If f ∈ C k (T) and g = f (k) , then gˆ(n) = (2πin)k f(n), n ∈ Z. Proof. Homework. Note: True under the weaker assumption that f ∈ C k−1 (T), g ∈ L1 (T), and Rt f k−1 (t) = f k−1 (0) + 0 g(s)ds.

31

CHAPTER 1. FINITE FOURIER TRANSFORM

1.4 1.4.1

Applications Wirtinger’s Inequality

Theorem 1.40 (Wirtinger’s Inequality). Suppose that f ∈ L2 (a, b), and that “f has a derivative in L2 (a, b)”, i.e., suppose that Z t f (t) = f (a) + g(s)ds a

where g ∈ L2 (a, b). In addition, suppose that f (a) = f (b) = 0. Then Z

a

b 2

|f (t)| dt ≤ =





b−a π b−a π

2 Z

b

|g(t)|2dt

a

2 Z

b

a

(1.8) !

|f ′(t)|2 dt .

Comment 1.41. A function f which can be written in the form Z t f (t) = f (a) + g(s)ds, a

where g ∈ L1 (a, b) is called absolutely continuous on (a, b). This is the “Lebesgue

version of differentiability”. See, for example, Rudin’s “Real and Complex Analysis”. Proof. i) First we reduce the interval (a, b) to (0, 1/2): Define F (s) = f (a + 2(b − a)s)

G(s) = F ′ (s) = 2(b − a)g(a + 2(b − a)s). Then F (0) = F (1/2) = 0 and F (t) =

Rt 0

G(s)ds. Change variable in the integral:

t = a + 2(b − a)s, dt = 2(b − a)ds, and (1.8) becomes Z

1/2

1 |F (s)| ds ≤ 2 4π 2

0

Z

0

1/2

|G(s)|2 ds.

(1.9)

We extend F and G to periodic functions, period one, so that F is odd and G is even: F (−t) = −F (t) and G(−t) = G(t) (first to the interval (−1/2, 1/2) and

32

CHAPTER 1. FINITE FOURIER TRANSFORM

then by periodicity to all of R). The extended function F is continuous since F (0) = F (1/2) = 0. Then (1.9) becomes Z Z 1 2 |F (s)| ds ≤ |G(s)|2 ds 2 4π T T 1 kF kL2 (T) ≤ kGkL2 (T) 2π

⇔

By Parseval’s identity, equation (1.6) on page 20, and Theorem 1.39 this is equivalent to

Here

∞ 1 X 2 ˆ |2πnFˆ (n)|2 . |F (n)| ≤ 2 4π n=−∞ n=−∞ ∞ X

Fˆ (0) =

Z

(1.10)

1/2

F (s)ds = 0.

−1/2

since F is odd, and for n 6= 0 we have (2πn)2 ≥ 4π 2 . Thus (1.10) is true. 
2 Note: The constant b−a is the best possible: we get equality if we take π Fˆ (1) 6= 0, Fˆ (−1) = −Fˆ (1), and all other Fˆ (n) = 0. (Which function is this?)

1.4.2

Weierstrass Approximation Theorem

Theorem 1.42 (Weierstrass Approximation Theorem). Every continuous function on a closed interval [a, b] can be uniformly approximated by a polynomial: For every ε > 0 there is a polynomial P so that max |P (t) − f (t)| ≤ ε

t∈[a,b]

(1.11)

Proof. First change the variable so that the interval becomes [0, 1/2] (see previous page). Then extend f to an even function on [−1/2, 1/2] (see previous page). Then extend f to a continuous 1-periodic function. By Corollary 1.35, the sequence fm (t) =

m X

Fˆm (n)fˆ(n)e2πint

n=−m

(Fm = Fejer kernel) converges to f uniformly. Choose m so large that |fm (t) − f (t)| ≤ ε/2 for all t. The function fm (t) is analytic, so by the course in analytic functions, the series

∞ (k) X fm (0) k=0

k!

tk

33

CHAPTER 1. FINITE FOURIER TRANSFORM

converges to fm (t), uniformly for t ∈ [−1/2, 1/2]. By taking N large enough we therefore have

|PN (t) − fm (t)| ≤ ε/2 for t ∈ [−1/2, 1/2], PN fm(k) (0) k where PN (t) = t . This is a polynomial, and |PN (t) − f (t)| ≤ ε k=0 k!

for t ∈ [−1/2, 1/2]. Changing the variable t back to the original one we get a polynomial satisfying (1.11).

1.4.3



Solution of Differential Equations

There are many ways to use Fourier series to solve differential equations. We give only two examples. Example 1.43. Solve the differential equation y ′′(x) + λy(x) = f (x),

0 ≤ x ≤ 1,

(1.12)

with boundary conditions y(0) = y(1), y ′(0) = y ′(1). (These are periodic boundary conditions.) The function f is given, and λ ∈ C is a constant. Solution. Extend y and f to all of R so that they become periodic, period 1. The equation + boundary conditions then give y ∈ C 1 (T). If we in addition

assume that f ∈ L2 (T), then (1.12) says that y ′′ = f − λy ∈ L2 (T) (i.e. f ′ is

“absolutely continuous”).

Assuming that f ∈ C 1 (T) and that f ′ is absolutely continuous we have by one of the homeworks

d ′′ )(n) = (2πin)2 y (y ˆ(n),

so by transforming (1.12) we get

−4π 2 n2 yˆ(n) + λˆ y (n) = fˆ(n),

(λ − 4π 2 n2 )ˆ y (n) = fˆ(n),

n ∈ Z, or

(1.13)

n ∈ Z.

Case A: λ 6= 4π 2 n2 for all n ∈ Z. Then (1.13) gives yˆ(n) =

fˆ(n) . λ − 4π 2 n2

The sequence on the right is in ℓ1 (Z), so yˆ(n) ∈ ℓ1 (Z). (i.e.,

Theorem 1.37,

y(t) =

∞ X

fˆ(n) e2πint , 2 n2 λ − 4π {z } n=−∞ | =ˆ y (n)

t ∈ R.

P |ˆ y (n)| < ∞). By

CHAPTER 1. FINITE FOURIER TRANSFORM

34

Thus, this is the only possible solution of (1.12). How do we know that it is, indeed, a solution? Actually, we don’t, but by working harder, and using the results from Chapter 0, it can be shown that y ∈ C 1 (T), and

y ′(t) =

∞ X

2πinˆ y (n)e2πint ,

n=−∞

where the sequence

2πinˆ y (n) = belongs to ℓ1 (Z) (both

2πin λ−4π 2 n2

2πinˆ y (n) λ − 4π 2 n2

and yˆ(n) belongs to ℓ2 (Z), and the product of two

ℓ2 -sequences is an ℓ1 -sequence; see Analysis II). The sequence (2πin)2 yˆ(n) =

−4π 2 n2 ˆ f(n) λ − 4π 2 n2

is an ℓ2 -sequence, and ∞ X

−4π 2 n2 ˆ f(n) → f ′′ (t) 2 2 λ − 4π n n=−∞ in the L2 -sense. Thus, f ∈ C 1 (T), f ′ is “absolutely continuous”, and equation (1.12) holds in the L2 -sense (but not necessary everywhere). (It is called a mild solution of (1.12)). Case B: λ = 4π 2 k 2 for some k ∈ Z. Write λ − 4π 2 n2 = 4π 2 (k 2 − n2 ) = 4π 2 (k − n)(k + n). ˆ We get two additional necessary conditions: f(±k) = 0. (If this condition is not true then the equation has no solutions.) If fˆ(k) = fˆ(−k) = 0, then we get infinitely many solutions: Choose yˆ(k) and yˆ(−k) arbitrarily, and yˆ(n) =

fˆ(n) , 4π 2 (k 2 − n2 )

n 6= ±k.

Continue as in Case A. Example 1.44. Same equation, but new boundary conditions: Interval is [0, 1/2], and y(0) = 0 = y(1/2).

CHAPTER 1. FINITE FOURIER TRANSFORM

35

Extend y and f to [−1/2, 1/2] as odd functions y(t) = −y(−t),

−1/2 ≤ t ≤ 0

f (t) = −f (−t),

−1/2 ≤ t ≤ 0

and then make them periodic, period 1. Continue as before. This leads to a Fourier series with odd coefficients, which can be rewritten as a sinus-series. Example 1.45. Same equation, interval [0, 1/2], boundary conditions y ′(0) = 0 = y ′ (1/2). Extend y and f to even functions, and continue as above. This leads to a solution with even coefficients yˆ(n), and it can be rewritten as a cosinus-series.

1.4.4

Solution of Partial Differential Equations

See course on special functions.

Chapter 2 Fourier Integrals 2.1

L1-Theory

Repetition: R = (−∞, ∞), 1

f ∈ L (R) ⇔ 2

f ∈ L (R) ⇔

Z

Z

∞ −∞ ∞

−∞

|f (t)|dt < ∞ (and f measurable)

|f (t)|2dt < ∞ (and f measurable)

Definition 2.1. The Fourier transform of f ∈ L1 (R) is given by Z ∞ ˆ f (ω) = e−2πiωt f (t)dt, ω ∈ R −∞

Comparison to chapter 1: ˆ f ∈ L1 (T) ⇒ f(n) defined for all n ∈ Z ˆ f ∈ L1 (R) ⇒ f(ω) defined for all ω ∈ R Notation 2.2. C0 (R) = “continuous functions f (t) satisfying f (t) → 0 as t → ±∞”. The norm in C0 is

kf kC0 (R) = max|f (t)| (= sup|f (t)|). t∈R

t∈R

Compare this to c0 (Z). Theorem 2.3. The Fourier transform F maps L1 (R) → C0 (R), and it is a ˆ C (R) ≤ kf kL1 (R) , i.e., contraction, i.e., if f ∈ L1 (R), then fˆ ∈ C0 (R) and kfk 0 36

CHAPTER 2. FOURIER INTEGRALS

37

i) fˆ is continuous ii) fˆ(ω) → 0 as ω → ±∞ iii) |fˆ(ω)| ≤

R∞

−∞

|f (t)|dt, ω ∈ R.

Note: Part ii) is again the Riemann-Lesbesgue lemma. Proof. iii) “The same” as the proof of Theorem 1.4 i). ii) “The same” as the proof of Theorem 1.4 ii), (replace n by ω, and prove this first in the special case where f is continuously differentiable and vanishes outside of some finite interval). i) (The only “new” thing): ˆ + h) − f(ω)| ˆ |f(ω

= = △-ineq.

≤





Z

ZR

Z

R

R


e−2πi(ω+h)t − e−2πiωt f (t)dt


e−2πiht − 1 e−2πiωt f (t)dt

|e−2πiht − 1||f (t)| dt → 0 as h → 0

(use Lesbesgue’s dominated convergens Theorem, e−2πiht → 1 as h → 0, and |e−2πiht − 1| ≤ 2).



Question 2.4. Is it possible to find a function f ∈ L1 (R) whose Fourier trans-

form is the same as the original function?

Answer: Yes, there are many. See course on special functions. All functions which are eigenfunctions with eigenvalue 1 are mapped onto themselves. Special case: 2 ˆ 0 (ω) = e−πω2 , ω ∈ R Example 2.5. If h0 (t) = e−πt , t ∈ R, then h

Proof. See course on special functions. Note: After rescaling, this becomes the normal (Gaussian) distribution function. This is no coincidence! Another useful Fourier transform is: Example 2.6. The Fejer kernel in L1 (R) is F (t) =

sin(πt)
2 . πt

38

CHAPTER 2. FOURIER INTEGRALS

The transform of this function is ( 1 − |ω| , Fˆ (ω) = 0 ,

|ω| ≤ 1,

otherwise.

Proof. Direct computation. (Compare this to the periodic Fejer kernel on page 23.) Theorem 2.7 (Basic rules). Let f ∈ L1 (R), τ, λ ∈ R ˆ ⇒ gˆ(ω) = e−2πiωτ f(ω) ⇒ gˆ(ω) = fˆ(ω − τ )

a) g(t) = f (t − τ )

b) g(t) = e2πiτ t f (t)

⇒ gˆ(ω) = fˆ(−ω) ⇒ gˆ(ω) = fˆ(−ω)

c) g(t) = f (−t) d) g(t) = f (t) e) g(t) = λf (λt) f ) g ∈ L1 and h = f ∗ g g) h)

g(t) = −2πitf (t)

and g ∈ L1

)

f is “absolutely continuous“ ′

1

and f = g ∈ L (R)

)

⇒ gˆ(ω) = fˆ( ωλ ) (λ > 0) ˆ ˆ g (ω) ⇒ h(ω) = f(ω)ˆ ( fˆ ∈ C 1 (R), and ⇒ ′ fˆ (ω) = gˆ(ω) ˆ ⇒ gˆ(ω) = 2πiω f(ω).

Proof. (a)-(e): Straightforward computation. (g)-(h): Homework(?) (or later). The formal inversion for Fourier integrals is Z ∞ ˆ f (ω) = e−2πiωt f (t)dt Z−∞ ∞ ? f (t) = e2πiωt fˆ(ω)dω −∞

This is true in “some cases” in “some sense”. To prove this we need some additional machinery. Definition 2.8. Let f ∈ L1 (R) and g ∈ Lp (R), where 1 ≤ p ≤ ∞. Then we define

(f ∗ g)(t) =

Z

R

f (t − s)g(s)ds

for all those t ∈ R for which this integral converges absolutely, i.e., Z |f (t − s)g(s)|ds < ∞. R

39

CHAPTER 2. FOURIER INTEGRALS

Lemma 2.9. With f and p as above, f ∗ g is defined a.e., f ∗ g ∈ Lp (R), and kf ∗ gkLp (R) ≤ kf kL1 (R) kgkLp (R) . If p = ∞, then f ∗ g is defined everywhere and uniformly continuous. Conclusion 2.10. If kf kL1 (R) ≤ 1, then the mapping g 7→ f ∗ g is a contraction from Lp (R) to itself (same as in periodic case).

Proof. p = 1: “same” proof as we gave on page 21. p = ∞: Boundedness of f ∗ g easy. To prove continuity we approximate f by a

function with compact support and show that kf (t) − f (t + h)kL1 → 0 as h → 0. p 6= 1, ∞: Significantly harder, case p = 2 found in Gasquet.

Notation 2.11. BUC(R) = “all bounded and continuous functions on R”. We

use the norm

kf kBU C(R) = sup|f (t)|. t∈R

ˆ Theorem 2.12 (“Approximate identity”). Let k ∈ L1 (R), k(0) =

1, and define

R∞

−∞

t ∈ R, λ > 0.

kλ (t) = λk(λt), If f belongs to one of the function spaces

a) f ∈ Lp (R), 1 ≤ p < ∞ (note: p 6= ∞), b) f ∈ C0 (R), c) f ∈ BUC(R), then kλ ∗ f belongs to the same function space, and kλ ∗ f → f

as λ → ∞

in the norm of the same function space, i.e., kkλ ∗ f − f kLp (R) → 0 as λ → ∞ if f ∈ Lp (R) ( if f ∈ BUC(R), supt∈R |(kλ ∗ f )(t) − f (t)| → 0 as λ → ∞ or f ∈ C0 (R). It also conveges a.e. if we assume that

R∞ 0

(sups≥|t| |k(s)|)dt < ∞.

k(t)dt =

40

CHAPTER 2. FOURIER INTEGRALS

Proof. “The same” as the proofs of Theorems 1.29, 1.32 and 1.33. That is, the computations stay the same, but the bounds of integration change (T → R), and the motivations change a little (but not much).  Example 2.13 (Standard choices of k). i) The Gaussian kernel 2 2 ˆ k(t) = e−πt , k(ω) = e−πω .

This function is C ∞ and nonnegative, so Z Z ˆ k(t)dt = k(0) = 1. kkkL1 = |k(t)|dt = R

R

ii) The Fejer kernel F (t) =

sin(πt)2 . (πt)2

It has the same advantages, and in addition Fˆ (ω) = 0 for |ω| > 1. The transform is a triangle: Fˆ (ω) =

(

1 − |ω|, |ω| ≤ 1 0,

|ω| > 1

F( ω)

−1

1

iii) k(t) = e−2|t| (or a rescaled version of this function. Here ˆ k(ω) =

1 , ω ∈ R. 1 + (πω)2

Same advantages (except C ∞ )).

41

CHAPTER 2. FOURIER INTEGRALS

ˆ Comment 2.14. According to Theorem 2.7 (e), kˆλ (ω) → k(0) = 1 as λ →

∞, for all ω ∈ R. All the kernels above are “low pass filters” (non causal).

It is possible to use “one-sided” (“causal”) filters instead (i.e., k(t) = 0 for t < 0). Substituting these into Theorem 2.12 we get “approximate identities”, which “converge to a δ-distribution”. Details later. Theorem 2.15. If both f ∈ L1 (R) and fˆ ∈ L1 (R), then the inversion formula Z ∞ f (t) = e2πiωt fˆ(ω)dω (2.1) −∞

is valid for almost all t ∈ R. By redefining f on a set of measure zero we can make it hold for all t ∈ R (the right hand side of (2.1) is continuous). Proof. We approximate R

R

R

e2πiωt fˆ(ω)dω by

2 2 (where ε > 0 is small) e2πiωt e−ε πω fˆ(ω)dω R 2πiωt−ε2 πω2 R −2πiωs = Re e f (s)dsdω (Fubini) R Z R 2 2 = s∈R f (s) e−2πiω(s−t) e|−ε{zπω} dωds (Ex. 2.13 last page)

R

ω∈R

k(εω 2 )

{z

|

(⋆)

}

(⋆) The Fourier transform of k(εω 2 ) at the point s − t. By Theorem 2.7 (e) this is equal to

1ˆ s − t )= = k( ε ε

1ˆ t − s k( ) ε ε

2 ˆ (since k(ω) = e−πω is even).

The whole thing is Z

1 f (s) k ε s∈R



t−s ε



ds = (f ∗ k 1 )(t) → f ∈ L1 (R) ε

as ε → 0+ according to Theorem 2.12. Thus, for almost all t ∈ R, Z 2 2 f (t) = lim e2πiωt e−ε πω fˆ(ω)dω. ε→0

R

On the other hand, by the Lebesgue dominated convergence theorem, since 2 2 |e2πiωt e−ε πω fˆ(ω)| ≤ |fˆ(ω)| ∈ L1 (R), Z Z 2πiωt −ε2 πω 2 ˆ ˆ e f (ω)dω = e2πiωt f(ω)dω. lim e

ε→0

R

R

42

CHAPTER 2. FOURIER INTEGRALS

Thus, (2.1) holds a.e. The proof of the fact that Z e2πiωt fˆ(ω)dω ∈ C0 (R) R

is the same as the proof of Theorem 2.3 (replace t by −t).



The same proof also gives us the following “approximate inversion formula”: Theorem 2.16. Suppose that k ∈ L1 (R), kˆ ∈ L1 (R), and that Z ˆ k(t)dt = 1. k(0) = R

If f belongs to one of the function spaces a) f ∈ Lp (R), 1 ≤ p < ∞ b) f ∈ C0 (R) c) f ∈ BUC(R) then

Z

R

ˆ ˆ e2πiωt k(εω) f(ω)dω → f (t)

in the norm of the given space (i.e., in Lp -norm, or in the sup-norm), and also R∞ a.e. if 0 (sups≥|t| |k(s)|)dt < ∞. Proof. Almost the same as the proof given above. If k is not even, then we

end up with a convolution with the function kε (t) = 1ε k(−t/ε) instead, but we can still apply Theorem 2.12 with k(t) replaced by k(−t).



Corollary 2.17. The inversion in Theorem 2.15 can be interpreted as follows: If f ∈ L1 (R) and fˆ ∈ L1 (R), then, ˆ fˆ(t) = f (−t) a.e. ˆ Here fˆ(t) = the Fourier transform of fˆ evaluated at the point t. Proof. By Theorem 2.15, f (t) =

Z

e−2πi(−t)ω fˆ(ω)dω |R {z }

Fourier transform of fˆ at the point (−t)

a.e.

43

CHAPTER 2. FOURIER INTEGRALS

ˆˆ ˆ Corollary 2.18. fˆ(t) = f (t) (If we repeat the Fourier transform 4 times, then we get back the original function). (True at least if f ∈ L1 (R) and fˆ ∈ L1 (R). ) As a prelude (=preludium) to the L2 -theory we still prove some additional results: Lemma 2.19. Let f ∈ L1 (R) and g ∈ L1 (R). Then Z Z ˆ f (t)ˆ g (t)dt = f(s)g(s)ds R

R

Proof. Z

f (t)ˆ g (t)dt =

R

= =

Z

f (t) t∈R Z Z Zs∈R

Z

e−2πits g(s)dsdt (Fubini) s∈R  −2πist f (t)e dt g(s)ds

t∈R

ˆ f(s)g(s)ds. 

s∈R

ˆ ∈ L1 (R). Then Theorem 2.20. Let f ∈ L1 (R), h ∈ L1 (R) and h Z Z ˆ f (t)h(t)dt = fˆ(ω)h(ω)dω. R

(2.2)

R

Specifically, if f = h, then (f ∈ L2 (R) and) kf kL2 (R) = kfˆkL2 (R) . R ˆ Proof. Since h(t) = ω∈R e2πiωt h(ω)dω we have Z Z Z ˆ f (t)h(t)dt = (Fubini) f (t) e−2πiωt h(ω)dωdt R t∈R ω∈R  Z Z −2πist ˆ = f (t)e dt h(ω)dω s∈R t∈R Z ˆ  = fˆ(ω)h(ω)dω. R

2.2

Rapidly Decaying Test Functions

(“Snabbt avtagande testfunktioner”). Definition 2.21. S = the set of functions f with the following properties i) f ∈ C ∞ (R) (infinitely many times differentiable)

(2.3)

44

CHAPTER 2. FOURIER INTEGRALS

ii) tk f (n) (t) → 0 as t → ±∞ and this is true for all k, n ∈ Z+ = {0, 1, 2, 3, . . . }. Thus: Every derivative of f → 0 at infinity faster than any negative power of t.

Note: There is no natural norm in this space (it is not a “Banach” space). However, it is possible to find a complete, shift-invariant metric on this space (it is a Frechet space). 2

Example 2.22. f (t) = P (t)e−πt ∈ S for every polynomial P (t). For example, the Hermite functions are of this type (see course in special functions).

Comment 2.23. Gripenberg denotes S by C↓∞ (R). The functions in S are called rapidly decaying test functions.

The main result of this section is Theorem 2.24. f ∈ S ⇐⇒ fˆ ∈ S That is, both the Fourier transform and the inverse Fourier transform maps this class of functions onto itself. Before proving this we prove the following Lemma 2.25. We can replace condition (ii) in the definition of the class S by one of the conditions R iii) R |tk f (n) (t)|dt < ∞, k, n ∈ Z+ or iv) |


d n k t f (t)| dt

→ 0 as t → ±∞, k, n ∈ Z+

without changing the class of functions S. Proof. If ii) holds, then for all k, n ∈ Z+ , sup|(1 + t2 )tk f (n) (t)| < ∞ t∈R

(replace k by k + 2 in ii). Thus, for some constant M, Z M k (n) =⇒ |tk f (n) (t)|dt < ∞. |t f (t)| ≤ 1 + t2 R Conversely, if iii) holds, then we can define g(t) = tk+1 f (n) (t) and get g ′(t) = (k + 1)tk f (n) (t) + tk+1f (n+1) (t), | {z } | {z } ∈L1

∈L1

45

CHAPTER 2. FOURIER INTEGRALS

so g ′ ∈ L1 (R), i.e.,

Z

∞

−∞

This implies

|g(t)| ≤ |g(0) +

Z

≤ |g(0)| + ≤ |g(0)| +

|g ′(t)|dt < ∞. t

g ′ (s)ds|

Z0

t

|g ′(s)|ds

Z0 ∞

−∞

|g ′(s)|ds = |g(0)| + kg ′kL1 ,

so g is bounded. Thus, 1 tk f (n) (t) = g(t) → 0 as t → ±∞. t The proof that ii) ⇐⇒ iv) is left as a homework.



Proof of Theorem 2.24. By Theorem 2.7, the Fourier transform of  k d k (n) ˆ (−2πit) f (t) is (2πiω)n f(ω). dω

Therefore, if f ∈ S, then condition iii) on the last page holds, and by Theorem 2.3, fˆ satisfies the condition iv) on the last page. Thus fˆ ∈ S. The same argument with e−2πiωt replaced by e+2πiωt shows that if fˆ ∈ S, then the Fourier inverse transform of fˆ (which is f ) belongs to S.



Note: Theorem 2.24 is the basis for the theory of Fourier transforms of distributions. More on this later.

2.3

L2-Theory for Fourier Integrals

As we saw earlier in Lemma 1.10, L2 (T) ⊂ L1 (T). However, it is not true that

L2 (R) ⊂ L1 (R). Counter example:

f (t) = √ (too large at ∞).

 2    ∈ L (R)

1 6 L1 (R) ∈ 1 + t2    ∈ C ∞ (R)

So how on earth should we define fˆ(ω) for f ∈ L2 (R), if the integral Z e−2πint f (t)dt R

CHAPTER 2. FOURIER INTEGRALS

46

does not converge? Recall: Lebesgue integral converges ⇐⇒ converges absolutely ⇐⇒ Z |e−2πint f (t)|dt < ∞ ⇐⇒ f ∈ L1 (R). We are saved by Theorem 2.20. Notice, in particular, condition (2.3) in that theorem! Definition 2.26 (L2 -Fourier transform). i) Approximate f ∈ L2 (R) by a sequence fn ∈ S which converges to f in

L2 (R). We do this e.g. by “smoothing” and “cutting” (“utj¨amning” och 2

“klippning”): Let k(t) = e−πt , define kn (t) = nk(nt), and   t (kn ∗ f )(t) fn (t) = k n | {z } | {z } ⋆⋆ ⋆ | {z }

the product belongs to S

(⋆) this tends to zero faster than any polynomial as t → ∞. (⋆⋆) “smoothing” by an approximate identity, belongs to C ∞ and is bounded.
By Theorem 2.12 kn ∗ f → f in L2 as n → ∞. The functions k nt tend to k(0) = 1 at every point t as n → ∞, and they are uniformly bounded by 1. By using the appropriate version of the Lesbesgue convergence we let fn → f in L2 (R) as n → ∞. ii) Since fn converges in L2 , also fˆn must converge to something in L2 . More about this in “Analysis II”. This follows from Theorem 2.20. (fn → f ⇒ fn Cauchy sequence ⇒ fˆn Cauchy seqence ⇒ fˆn converges.) iii) Call the limit to which fn converges “The Fourier transform of f ”, and denote it fˆ. Definition 2.27 (Inverse Fourier transform). We do exactly as above, but replace e−2πiωt by e+2πiωt . Final conclusion:

47

CHAPTER 2. FOURIER INTEGRALS

Theorem 2.28. The “extended” Fourier transform which we have defined above has the following properties: It maps L2 (R) one-to-one onto L2 (R), and if fˆ is the Fourier transform of f , then f is the inverse Fourier transform of fˆ. Moreover, all norms, distances and inner products are preserved. Explanation: i) “Normes preserved” means Z Z 2 |f (t)| dt = |fˆ(ω)|2dω, R

R

ˆ L2 (R) . or equivalently, kf kL2 (R) = kfk

ii) “Distances preserved” means kf − gkL2(R) = kfˆ − gˆkL2 (R) (apply i) with f replaced by f − g) iii) “Inner product preserved” means Z Z g (ω)dω, fˆ(ω)ˆ f (t)g(t)dt = R

R

which is often written as

hf, giL2(R) = hfˆ, gˆiL2 (R) . This was theory. How to do in practice? One answer: We saw earlier that if [a, b] is a finite interval, and if f ∈ L2 [a, b] ⇒

f ∈ L1 [a, b], so for each T > 0, the integral Z T ˆ fT (ω) = e−2πiωt f (t)dt −T

is defined for all ω ∈ R. We can try to let T → ∞, and see what happens. (This resembles the theory for the inversion formula for the periodical L2 -theory.) Theorem 2.29. Suppose that f ∈ L2 (R). Then Z T lim e−2πiωt f (t)dt = fˆ(ω) T →∞

−T

2

in the L -sense as T → ∞, and likewise Z T lim e2πiωt fˆ(ω)dω = f (t) T →∞

in the L2 -sense.

−T

48

CHAPTER 2. FOURIER INTEGRALS

Proof. Much too hard to be presented here. Another possibility: Use the Fejer kernel or the Gaussian kernel, or any other kernel, and define
R ˆ f(ω) = limn→∞ R e−2πiωt k nt f (t)dt, R
f (t) = limn→∞ e+2πiωt kˆ ω fˆ(ω)dω. n

R

We typically have the same type of convergence as we had in the Fourier inversion formula in the periodic case. (This is a well-developed part of mathematics, with lots of results available.) See Gripenberg’s compendium for some additional results.

2.4

An Inversion Theorem

From time to time we need a better (= more useful) inversion theorem for the Fourier transform, so let us prove one here: Theorem 2.30. Suppose that f ∈ L1 (R) + L2 (R) (i.e., f = f1 + f2 , where f1 ∈ L1 (R) and f2 ∈ L2 (R)). Let t0 ∈ R, and suppose that Z t0 +1 f (t) − f (t0 ) dt < ∞. t − t0 t0 −1

Then

f (t0 ) = lim ˆ where f(ω) = fˆ1 (ω) + fˆ2 (ω).

S→∞ T →∞

Z

T

e2πiωt0 fˆ(ω)dω,

(2.4)

(2.5)

−S

Comment: Condition (2.4) is true if, for example, f is differentiable at the point t0 . Proof. Step 1. First replace f (t) by g(t) = f (t + t0 ). Then ˆ gˆ(ω) = e2πiωt0 f(ω), and (2.5) becomes g(0) = lim

S→∞ T →∞

and (2.4) becomes

Z

T

gˆ(ω)dω, −S

Z 1 g(t − t0 ) − g(0) dt < ∞. t − t0 −1

Thus, it suffices to prove the case where t0 = 0 .

49

CHAPTER 2. FOURIER INTEGRALS

2

Step 2: We know that the theorem is true if g(t) = e−πt (See Example 2.5 and Theorem 2.15). Replace g(t) by 2

h(t) = g(t) − g(0)e−πt . Then h satisfies all the assumptions which g does, and in addition, h(0) = 0. Thus it suffices to prove the case where both (⋆) t0 = 0 and f (0) = 0 . For simplicity we write f instead of h but assume (⋆). Then (2.4) and (2.5) simplify: Z 1 f (t) dt < ∞, t −1 Z T ˆ lim f(ω)dω = 0.

S→∞ T →∞

(2.6) (2.7)

−S

Step 3: If f ∈ L1 (R), then we argue as follows. Define g(t) =

f (t) . −2πit

Then g ∈ L1 (R). By Fubini’s theorem, Z T Z TZ ∞ ˆ f (ω)dω = e−2πiωt f (t)dtdω −S −S −∞ Z ∞Z T = e−2πiωt dωf (t)dt = =

−∞ ∞

Z

−∞ Z ∞ −∞





−S

1 e−2πiωt −2πit

T

f (t)dt

−S

e−2πiT t − e−2πi(−S)t

= gˆ(T ) − gˆ(−S),

 f (t) dt −2πit

and this tends to zero as T → ∞ and S → ∞ (see Theorem 2.3). This proves

(2.7).

Step 4: If instead f ∈ L2 (R), then we use Parseval’s identity Z ∞ Z ∞ ˆ h(ω)dω ˆ f(ω) f (t)h(t)dt = −∞

−∞

in a clever way: Choose ˆ h(ω) =

(

1, −S ≤ t ≤ T,

0, otherwise.

50

CHAPTER 2. FOURIER INTEGRALS

ˆ is Then the inverse Fourier transform h(t) of h Z T h(t) = e2πiωt dω =



−S

1 2πiωt e 2πit

so Parseval’s identity gives Z T Z ˆ f (ω)dω =

∞

T

=

−S

 1  2πiT t e − e2πi(−S)t 2πit

 1  −2πiT t e − e−2πi(−S)t dt −2πit −∞ = (with g(t) as in Step 3) Z ∞  −2πiT t  = e − e−2πi(S)t g(t)dt −∞ ( T → ∞, = gˆ(T ) − gˆ(−S) → 0 as S → ∞.

−S

f (t)

Step 5: If f = f1 + f2 , where f1 ∈ L1 (R) and f2 ∈ L2 (R), then we apply Step 3

to f1 and Step 4 to f2 , and get in both cases (2.7) with f replaced by f1 and f2 .  Note: This means that in “most cases” where f is continuous at t0 we have Z T e2πiωt0 fˆ(ω)dω. f (t0 ) = lim S→∞ T →∞

−S

(continuous functions which do not satisfy (2.4) do exist, but they are difficult to find.) In some cases we can even use the inversion formula at a point where f is discontinuous. Theorem 2.31. Suppose that f ∈ L1 (R) + L2 (R). Let t0 ∈ R, and suppose that

the two limits

f (t0 +) = lim f (t) t↓t0

f (t0 −) = lim f (t) t↑t0

exist, and that Z

t0 +1

f (t) − f (t0 +) dt < ∞, t − t0 t0 Z t0 f (t) − f (t0 −) dt < ∞. t − t0 t0 −1

51

CHAPTER 2. FOURIER INTEGRALS

Then

Z

T

1 ˆ e2πiωt0 f(ω)dω = [f (t0 +) + f (t0 −)]. T →∞ −T 2 RT RT Note: Here we integrate −T , not −S , and the result is the average of the right lim

and left hand limits.

Proof. As in the proof of Theorem 2.30 we may assume that Step 1: t0 = 0 , (see Step 1 of that proof) Step 2: f (t0 +) + f (t0 −) = 0 , (see Step 2 of that proof). Step 3: The claim is true in the special case where ( e−t , t > 0, g(t) = −et , t < 0,

because g(0+) = 1, g(0−) = −1, g(0+) + g(0−) = 0, and Z T gˆ(ω)dω = 0 for all T, −T

since f is odd =⇒ gˆ is odd.

Step 4: Define h(t) = f (t) − f (0+) · g(t), where g is the function in Step 3. Then h(0+) = f (0+) − f (0+) = 0 and h(0−) = f (0−) − f (0+)(−1) = 0,

so

h is continuous. Now apply Theorem 2.30 to h. It gives Z T ˆ 0 = h(0) = lim h(ω)dω. T →∞

Since also

−T

0 = f (0+)[g(0+) + g(0−)] = lim

T →∞

we therefore get 0 = f (0+) + f (0−) = lim

T →∞

Z

T

Z

T

gˆ(ω)dω,

−T

ˆ [h(ω) + gˆ(ω)]dω = lim

T →∞

−T

Z

T

ˆ f(ω)dω. 

−T

Comment 2.32. Theorems 2.30 and 2.31 also remain true if we replace Z T lim e2πiωt fˆ(ω)dω T →∞

by

lim

ε→0

Z

∞

−T

2 ˆ e2πiωt e−π(εω) f(ω)dω

(2.8)

−∞

(and other similar “summability” formulas). Compare this to Theorem 2.16. In 2

the case of Theorem 2.31 it is important that the “cutoff kernel” (= e−π(εω) in (2.8)) is even.

52

CHAPTER 2. FOURIER INTEGRALS

2.5 2.5.1

Applications The Poisson Summation Formula

P 1 ˆ ˆ Suppose that f ∈ L1 (R) ∩ C(R), that ∞ n=−∞ |f (n)| < ∞ (i.e., f ∈ ℓ (Z)), and P that ∞ n=−∞ f (t + n) converges uniformly for all t in some interval (−δ, δ). Then ∞ X

f (n) =

n=−∞

∞ X

fˆ(n)

(2.9)

n=−∞

P

Note: The uniform convergence of possible way out is: If we define

f (t + n) can be difficult to check. One

εn = sup |f (t + n)|, and if

P∞

n=−∞

εn < ∞, then

P∞

−δ
f (t + n) converges (even absolutely), and

n=−∞

the convergence is uniform in (−δ, δ). The proof is roughly the same as what we did on page 29. We first construct a periodic function g ∈ L1 (T) with the Fourier coefficients fˆ(n): Z ∞ ˆ f (n) = e−2πint f (t)dt

Proof of (2.9).

= t=k+s

=

−∞ ∞ X

Z

=

=

e−2πint f (t)dt

k=−∞ k ∞ Z 1 X k=−∞

Thm 0.14

k+1

Z

e−2πins f (s + k)ds

0

1

−2πins

e 0

gˆ(n),

∞ X

!

f (s + k) ds

k=−∞

where g(t) =

∞ X

f (t + n).

n=−∞

(For this part of the proof it is enough to have f ∈ L1 (R). The other conditions

are needed later.)

So we have gˆ(n) = fˆ(n). By the inversion formula for the periodic Fourier transform: g(0) =

∞ X

n=−∞

e2πin0 gˆ(n) =

∞ X

n=−∞

gˆ(n) =

∞ X

n=−∞

fˆ(n),

53

CHAPTER 2. FOURIER INTEGRALS

provided (=f¨orutsatt) that we are allowed to use the Fourier inversion formula. This is allowed if g ∈ C[−δ, δ] and gˆ ∈ ℓ1 (Z) (Theorem 1.37). This was part of our assumption.

In addition we need to know that the formula ∞ X g(t) = f (t + n) n=−∞

holds at the point t = 0 (almost everywhere is no good, we need it in exactly P this point). This is OK if ∞ n=−∞ f (t + n) converges uniformly in [−δ, δ] (this also implies that the limit function g is continuous).

Note: By working harder in the proof, Gripenberg is able to weaken some of the assumptions. There are also some counter-examples on how things can go wrong if you try to weaken the assumptions in the wrong way.

2.5.2

\ 1 (R) = C (R) ? Is L 0

That is, is every function g ∈ C0 (R) the Fourier transform of a function f ∈ L1 (R)?

The answer is no, as the following counter-example shows. Take  ω  , |ω| ≤ 1,  ln 2  1 g(ω) = , ω > 1, ln(1+ω)    − 1 , ω < −1. ln(1−ω)

Suppose that this would be the Fourier transform of a function f ∈ L1 (R). As in the proof on the previous page, we define h(t) =

∞ X

f (t + n).

n=−∞

ˆ Then (as we saw there), h ∈ L1 (T), and h(n) = fˆ(n) for n = 0, ±1, ±2, . . .. P∞ 1 ˆ However, since n=1 n h(n) = ∞, this is not the Fourier sequence of any h ∈ L1 (T) (by Theorem 1.38). Thus:

Not every h ∈ C0 (R) is the Fourier transform of some f ∈ L1 (R). But: f ∈ L1 (R) ⇒ fˆ ∈ C0 (R) ( Page 36) f ∈ L2 (R) ⇔ fˆ ∈ L2 (R) ( Page 47) f ∈S ⇔ fˆ ∈ S ( Page 44)

54

CHAPTER 2. FOURIER INTEGRALS

2.5.3

The Euler-MacLauren Summation Formula

Let f ∈ C ∞ (R+ ) (where R+ = [0, ∞)), and suppose that f (n) ∈ L1 (R+ ) for all n ∈ Z+ = {0, 1, 2, 3 . . .}. We define f (t) for t < 0 so that f (t) is even.

Warning: f is continuous at the origin, but f ′ may be discontinuous! For exam-

ple, f (t) = e−|2t|

f(t)=e −2|t|

We want to use Poisson summation formula. Is this allowed? (n) = (2πiω)n f(ω), ˆ By Theorem 2.7, fd and fˆ(n) is bounded, so ˆ sup|(2πiω) ||f(ω)| < ∞ for all n ⇒ n

ω∈R

By the note on page 52, also

P∞

n=−∞

∞ X

ˆ |f(n)| < ∞.

n=−∞

f (t + n) converges uniformly in (−1, 1). By

the Poisson summation formula: ∞ X

∞ 1 1 X f (n) = f (0) + f (n) 2 2 n=0 n=−∞ ∞ 1 X ˆ 1 f(n) f (0) + = 2 2 n=−∞

i 1 1 1 Xhˆ ˆ = f (0) + fˆ(0) + f(n) + f(−n) 2 2 2 n=1 ∞ Z ∞ X
1ˆ 1 2πint 1 e + e−2πint f (t)dt f (0) + f (0) + = 2 2 2 {z } n=1 −∞ | ∞

=

1 f (0) + 2

Z

0

∞

f (t)dt +

∞ Z ∞ X n=1

cos(2πnt)

cos(2πnt)f (t)dt

0

Here we integrate by parts several times, always integrating the cosine-function and differentiating f . All the substitution terms containing odd derivatives of

55

CHAPTER 2. FOURIER INTEGRALS

f vanish since sin(2πnt) = 0 for t = 0. See Gripenberg for details. The result looks something like ∞ X

f (n) =

n=0

2.5.4

Z

∞ 0

1 1 1 ′′′ 1 f (t)dt + f (0) − f ′ (0) + f (0) − f (5) (0) + . . . 2 12 720 30240

Schwartz inequality

The Schwartz inequality will be used below. It says that |hf, gi| ≤ kf kL2 kgkL2 (true for all possible L2 -spaces, both L2 (R) and L2 (T) etc.)

2.5.5

Heisenberg’s Uncertainty Principle

For all f ∈ L2 (R), we have Z ∞  Z 2 2 t |f (t)| dt −∞

∞

−∞

ω |fˆ(ω)|2dω 2



≥

1

f 4 2 2 L (R) 16π

Interpretation: The more concentrated f is in the neighborhood of zero, the more spread out must fˆ be, and conversely. (Here we must think that kf kL2 (R) is fixed, e.g. kf kL2 (R) = 1.)

In quantum mechanics: The product of “time uncertainty” and “space uncer-

tainty” cannot be less than a given fixed number.

56

CHAPTER 2. FOURIER INTEGRALS

Proof. We begin with the case where f ∈ S. Then Z Z Z Z ˆ 16π |tf (t)|dt |ω f(ω)|dω = 4 |tf (t)|dt |f ′ (t)|dt R

R

R

R

′ (ω) = 2πiω f(ω) ˆ (f[ and Parseval’s iden. holds). Now use Scwartz ineq. Z  ′ ≥ 4 |tf (t)||f (t)|dt R Z  ′ = 4 |tf (t)||f (t)|dt R Z  ′ Re[tf (t)f (t)]dt ≥ 4 R Z    2 1 ′ ′ f (t)f (t) + f (t)f (t) dt = 4 t 2 R Z d t (f (t)f (t)) dt (integrate by parts) = {z } R dt | =|f (t)|

=

=

Z ∞   ∞ |f (t)|dt [t|f (t)|]−∞ − | {z } −∞ =0 Z ∞  |f (t)|dt −∞

This proves the case where f ∈ S. If f ∈ L(R), but f ∈ S, then we choose a sequence of functions fn ∈ S so that Z ∞ Z ∞ |fn (t)|dt → |f (t)|dt and −∞ −∞ Z ∞ Z ∞ |tfn (t)|dt → |tf (t)|dt and −∞ −∞ Z ∞ Z ∞ ˆ ˆ |ω fn (ω)|dω → |ω f(ω)|dω −∞

−∞

(This can be done, not quite obvious). Since the inequality holds for each fn , it must also hold for f .

2.5.6

Weierstrass’ Non-Differentiable Function

Define σ(t) =

P∞

k k=0 a

cos(2πbk t), t ∈ R where 0 < a < 1 and ab≥ 1.

Lemma 2.33. This sum defines a continuous function σ which is not differentiable at any point.

57

CHAPTER 2. FOURIER INTEGRALS

Proof. Convergence easy: At each t, ∞ X k=0

k

k

|a cos(2πb t)| ≤

∞ X

ak =

k=0

1 < ∞, 1−a

and absolute convergence ⇒ convergence. The convergence is even uniform: The

error is

∞ ∞ ∞ X X X aK k k k k → 0 as K → ∞ ak = |a cos(2πb t)| ≤ a cos(2πb t) ≤ 1−a k=K

k=K

k=K

so by choosing K large enough we can make the error smaller than ε, and the same K works for all t. By “Analysis II”: If a sequence of continuous functions converges uniformly, then the limit function is continuous. Thus, σ is continuous. Why is it not differentiable? At least does the formal derivative not converge: Formally we should have ′

σ (t) =

∞ X k=0

ak · 2πbk (−1) sin(2πbk t),

and the terms in this serie do not seem to go to zero (since (ab)k ≥ 1). (If a sum converges, then the terms must tend to zero.)

To prove that σ is not differentiable we cut the sum appropriatly: Choose some function ϕ ∈ L1 (R) with the following properties: i) ϕ(1) ˆ =1 ii) ϕ(ω) ˆ = 0 for ω ≤ iii)

R∞

−∞

1 b

and ω > b

|tϕ(t)|dt < ∞.

ϕ(ω) 0

1/b

1

b

We can get such a function from the Fejer kernel: Take the square of the Fejer kernel (⇒ its Fourier transform is the convolution of fˆ with itself), squeeze it (Theorem 2.7(e)), and shift it (Theorem 2.7(b)) so that it vanishes outside of

58

CHAPTER 2. FOURIER INTEGRALS

( 1b , b), and ϕ(1) ˆ = 1. (Sort of like approximate identity, but ϕ(1) ˆ = 1 instead of ϕ(0) ˆ = 1.) Define ϕj (t) = bj ϕ(bj t),

t ∈ R. Then ϕˆj (ω) = ϕ(ωb ˆ −j ), so ϕ(b ˆ j ) = 1 and

ϕ(ω) ˆ = 0 outside of the interval (bj−1 , bj+1 ).

ϕ3

b4

Put fj = σ ∗ ϕj . Then fj (t) = =

Z

∞

b3

ϕ2

b2

ϕ1

b

ϕ0

1

1/b

σ(t − s)ϕj (s)ds

−∞ Z ∞X ∞

k1

a

2

−∞ k=0

i h 2πibk (t−s) −2πibk (t−s) ϕj (s)ds e +e

(by the uniform convergence)   ∞ k Xa   2πibk t k −2πibk t = ϕj (−bk ) |e {z } ϕj (b ) + e {z } | 2 k=0

=

=δjk

=0

1 j 2πibk t ae . 2

(Thus, this particular convolution picks out just one of the terms in the series.) Suppose (to get a contradiction) that σ can be differentiated at some point t ∈ R.

Then the function

η(s) =

(

σ(t+s)−σ(t) s

0

− σ ′ (t) , s 6= 0 , s=0

is (uniformly) continuous and bounded, and η(0) = 0. Write this as σ(t − s) = −sη(−s) + σ(t) − sσ ′ (t)

59

CHAPTER 2. FOURIER INTEGRALS

i.e., fj (t)

= =

Z

ZR R

σ(t − s)ϕj (s)ds −sη(−s)ϕj (s)ds + σ(t)

Z

′

ϕj (s)ds −σ (t) sϕj (s)ds | R {z } | R {z } =ϕ ˆj (0)=0

= bj s=t

=

→0

Z

sη(−s)bj ϕ(bj s)ds RZ −s η( j ) sϕ(s)ds −bj b } | {z } R | {z ∈L1

−

Z

ϕ ˆ′ (0) j =0 −2πi

→0 pointwise

by the Lesbesgue dominated convergence theorem as j → ∞.

Thus,

1  a j 2πibj t b fj (t) → 0 as j → ∞ ⇔ e → 0 as j → ∞. 2 b
j j As |e2πib t | = 1, this is ⇔ ab → 0 as j → ∞. Impossible, since ab ≥ 1. Our −j

assumption that σ is differentiable at the point t must be wrong ⇒ σ(t) is not differentiable in any point!

2.5.7

Differential Equations

Solve the differential equation u′′(t) + λu(t) = f (t), t ∈ R

(2.10)

where we require that f ∈ L2 (R), u ∈ L2 (R), u ∈ C 1 (R), u′ ∈ L2 (R) and that u′ is of the form

′

′

u (t) = u (0) +

Z

t

v(s)ds, 0

where v ∈ L2 (R) (that is, u′ is “absolutely continuous” and its “generalized derivative” belongs to L2 ).

The solution of this problem is based on the following lemmas: Lemma 2.34. Let k = 1, 2, 3, . . .. Then the following conditions are equivalent: i) u ∈ L2 (R)∩C k−1 (R), u(k−1) is “absolutely continuous” and the “generalized derivative of u(k−1) ” belongs to L2 (R).

60

CHAPTER 2. FOURIER INTEGRALS

ii) uˆ ∈ L2 (R) and

R

R

|ω k uˆ(k)|2 dω < ∞.

Proof. Similar to the proof of one of the homeworks, which says that the same P ˆ 2 result is true for L2 -Fourier series. (There ii) is replaced by |nf(n)| < ∞.) Lemma 2.35. If u is as in Lemma 2.34, then

d (k) (ω) = (2πiω)k u u ˆ(ω)

(compare this to Theorem 2.7(g)).

Proof. Similar to the same homework. Solution: By the two preceding lemmas, we can take Fourier transforms in (2.10), and get the equivalent equation (2πiω)2 uˆ(ω)+λˆ u(ω) = fˆ(ω), ω ∈ R ⇔ (λ−4π 2 ω 2)ˆ u(ω) = fˆ(ω), ω ∈ R (2.11) Two cases: Case 1: λ − 4π 2 ω 2 6= 0, for all ω ∈ R, i.e., λ must not be zero and not a positive

number (negative is OK, complex is OK). Then uˆ(ω) =

fˆ(ω) , ω∈R λ − 4π 2 ω 2

so u = k ∗ f , where k = the inverse Fourier transform of ˆ k(ω) =

1 . λ − 4π 2 ω 2

This can be computed explicitly. It is called “Green’s function” for this problem. Even without computing k(t), we know that • k ∈ C0 (R) (since kˆ ∈ L1 (R).) • k has a generalized derivative in L2 (R) (since

R

R

2 ˆ |ω k(ω)| dω < ∞.)

• k does not have a second generalized derivative in L2 (since

R

R

2 ˆ |ω 2 k(ω)| dω = ∞.)

How to compute k? Start with a partial fraction expansion. Write λ = α2

for some α ∈ C

61

CHAPTER 2. FOURIER INTEGRALS

(α = pure imaginary if λ < 0). Then 1 λ − 4π 2 ω 2

Now we must still invert

1 1 1 = · α2 − 4π 2 ω 2 α − 2πω α + 2πω A B = + α − 2πω α + 2πω Aα + 2πωA + Bα − 2πωB = (α − 2πω)(α + 2πω) ) (A + B)α = 1 1 ⇒A=B= ⇒ 2α (A − B)2πω = 0 =

1 α+2πω

and

1 . α−2πω

This we do as follows:

Auxiliary result 1: Compute the transform of ( e−zt , t ≥ 0, f (t) = 0 , t < 0, / C(R) because of the jump at where Re(z) > 0 (⇒ f ∈ L2 (R) ∩ L1 (R), but f ∈

the origin). Simply compute: ˆ f(ω) =

Z

∞

e−2πiωt e−zt dt  0 −(z+2πiω)t ∞ e 1 = = . −(z + 2πiω) 0 2πiω + z

Auxiliary result 2: Compute the transform of ( ezt , t ≤ 0, f (t) = 0 , t > 0, where Re(z) > 0 (⇒ f ∈ L2 (R) ∩ L1 (R), but f ∈ / C(R)) ⇒ fˆ(ω) = = Back to the function k:

Z 

0

e2πiωt ezt dt −∞

e(z−2πiω)t (z − 2πiω)t

0

−∞

=

1 . z − 2πiω

  1 1 1 + 2α α − 2πω α + 2πω   i i 1 . + = 2α iα − 2πiω iα + 2πiω

ˆ k(ω) =

62

CHAPTER 2. FOURIER INTEGRALS

We defined α by requiring α2 = λ. Choose α so that Im(α) < 0 (possible because α is not a positive real number). 1 ˆ ⇒ Re(iα) > 0, and k(ω) = 2α



i i + iα − 2πiω iα + 2πiω



The auxiliary results 1 and 2 gives: k(t) =

(

i −iαt e 2α i iαt e 2α

and u(t) = (k ∗ f )(t) =

Z

, t≥0 , t<0

∞

−∞

k(t − s)f (s)ds

Special case: λ = negative number = −a2 , where a > 0. Take α = −ia

⇒ iα = i(−i)a = a, and

k(t) =

(

1 −at e , t≥0 − 2a

1 at − 2a e

k(t) = −

, t<0

i.e.

1 −|at| e , t∈R 2a

Thus, the solution of the equation u′′(t) − a2 u(t) = f (t),

t ∈ R,

where a > 0, is given by u=k∗f k(t) = −

where

1 −a|t| e , 2a

t∈R

This function k has many names, depending on the field of mathematics you are working in: i) Green’s function (PDE-people) ii) Fundamental solution (PDE-people, Functional Analysis) iii) Resolvent (Integral equations people)

63

CHAPTER 2. FOURIER INTEGRALS

Case 2: λ = a2 = a nonnegative number. Then fˆ(ω) = (a2 − 4π 2 ω 2 )ˆ u(ω) = (a − 2πω)(a + 2πω)ˆ u(ω). As uˆ(ω) ∈ L2 (R) we get a necessary condition for the existence of a solution: If a solution exists then

Z

R



2 fˆ(ω) dω < ∞. (a − 2πω)(a + 2πω)

(2.12)

a a , this forces fˆ to vanish at ± 2π , (Since the denominator vanishes for ω = ± 2π

and to be “small” near these points.)

If the condition (2.12) holds, then we can continue the solution as before. Sideremark: These results mean that this particular problem has no “eigenvalues” and no “eigenfunctions”. Instead it has a “contionuous spectrum” consisting of the positive real line. (Ignore this comment!)

2.5.8

Heat equation

This equation:    

∂ u(t, x) ∂t

   u(0, x)

=

∂2 ∂x2

u(t, x) + g(t, x),

(

= f (x) (initial value)

t>0 x∈R

is solved in the same way. Rather than proving everything we proceed in a formal mannor (everything can be proved, but it takes a lot of time and energy.) Transform the equation in the x-direction, Z uˆ(t, γ) = e−2πiγx u(t, x)dx. R

Assuming that

R

R

∂ e−2πiγx ∂t u(t, x) =

( (

∂ uˆ(t, γ) ∂t

uˆ(0, γ) ∂ uˆ(t, γ) ∂t

uˆ(0, γ)

∂ ∂t

R

R

e−2πiγx u(t, x)dx we get

= (2πiγ)2 uˆ(t, γ) + gˆ(t, γ) ˆ = f(γ) ⇔

= −4π 2 γ 2 uˆ(t, γ) + gˆ(t, γ) ˆ = f(γ)

64

CHAPTER 2. FOURIER INTEGRALS

We solve this by using the standard “variation of constants lemma”: 2 2 uˆ(t, γ) = fˆ(γ)e−4π γ t + | {z }

=

We can invert e−4π

2γ2t

uˆ1(t, γ) √

= e−π(2

+

πtγ)2

Z

|0

t

e−4π

2

= e−π(γ/λ)

2 γ 2 (t−s)

gˆ(s, γ)ds {z } uˆ2 (t, γ) √ where λ = (2 πt)−1 :

According to Theorem 2.7 and Example 2.5, this is the transform of x2 1 1 −π( √x )2 k(t, x) = √ e 2 πt = √ e− 4t 2 πt 2 πt

ˆ We know that fˆ(γ)k(γ) = k[ ∗ f (γ), so u1(t, x) =

R∞

2 √1 e−(x−y) /4t f (y)dy, −∞ 2 πt

(By the same argument:

s and t − s are fixed when we transform.) Rt u2(t, x) = 0 (k ∗ g)(s)ds RtR∞ 2 = 0 −∞ √ 1 e−(x−y) /4(t−s) g(s, y)dyds, 2

π(t−s)

u(t, x) = u1 (t, x) + u2 (t, x)

The function

x2 1 k(t, x) = √ e− 4t 2 πt is the Green’s function or the fundamental solution of the heat equation on the

real line R = (−∞, ∞), or the heat kernel.

Note: To prove that this “solution” is indeed a solution we need to assume that - all functions are in L2 (R) with respect to x, i.e., Z ∞ Z ∞ Z 2 2 |u(t, x)| dx < ∞, |g(t, x)| dx < ∞, −∞

−∞

∞

−∞

|f (x)|2 dx < ∞,

- some (weak) continuity assumptions with respect to t.

2.5.9

Wave equation           

∂2 ∂t2

u(t, x) =

∂2 ∂x2

u(t, x) + k(t, x),

u(0, x) = f (x), ∂ u(0, x) ∂t

= g(x),

(

t > 0, x ∈ R.

x∈R x∈R

65

CHAPTER 2. FOURIER INTEGRALS

Again we proceed formally. As  ∂2    ∂t2 uˆ(t, γ) uˆ(0, γ)    ∂ uˆ(0, γ)

above we get ˆ γ), = −4π 2 γ 2 uˆ(t, γ) + k(t, ˆ = f(γ), = gˆ(γ).

∂t

This can be solved by “the variation of constants formula”, but to simplify the ˆ γ) ≡ 0. Then the solution is computations we assume that k(t, x) ≡ 0, i.e., h(t, (check this!)

sin(2πγt) uˆ(t, γ) = cos(2πγt)fˆ(γ) + gˆ(γ). 2πγ

(2.13)

To invert the first term we use Theorem 2.7, and get 1 [f (x + t) + f (x − t)]. 2 The second term contains the “Dirichlet kernel ”, which is inverted as follows: Ex. If k(x) = ˆ then k(ω) = Proof.

1 2πω

(

1/2, |t| ≤ 1

0,

otherwise,

sin(2πω). 1 ˆ k(ω) = 2

Z

1

e−2πiωt dt = . . . = −1

1 sin(ωt). 2πω

Thus, the inverse Fourier transform of sin(2πγ) 2πγ

is k(x) =

(

1/2, |x| ≤ 1,

0,

|x| > 1,

(inverse transform = ordinary transform since the function is even), and the inverse Fourier transform (with respect to γ) of sin(2πγt) sin(2πγt) = t is 2πγ 2πγt ( 1/2, |x| ≤ t, x k( ) = t 0, |x| > t. This and Theorem 2.7(f), gives the inverse of the second term in (2.13): It is Z 1 x+t g(y)dy. 2 x−t

66

CHAPTER 2. FOURIER INTEGRALS

Conclusion: The solution of the wave equation with h(t, x) ≡ 0 seems to be 1 1 u(t, x) = [f (x + t) + f (x − t)] + 2 2

Z

x+t

g(y)dy,

x−t

a formula known as d’Alembert’s formula. Interpretation: This is the sum of two waves: u(t, x) = u+ (t, x) + u− (t, x), where 1 1 u+ (t, x) = f (x + t) + G(x + t) 2 2 moves to the left with speed one, and 1 1 u− (t, x) = f (x − t) − G(x − t) 2 2 moves to the right with speed one. Here Z x G(x) = g(y)dy, 0

x ∈ R.

Chapter 3 Fourier Transforms of Distributions Questions 1) How do we transform a function f ∈ / L1 (R), f ∈ / L2 (R), for example Weierstrass function

σ(t) =

∞ X

ak cos(2πbk t),

k=0

where b 6= integer (if b is an integer, then σ is periodic and we can use

Chapter I)?

2) Can we interpret both the periodic F -transform (on L1 (T)) and the Fourier integral (on L1 (R)) as special cases of a “more general” Fourier transform?

3) How do you differentiate a discontinuous function? The answer: Use “distribution theory”, developed in France by Schwartz in 1950’s.

3.1

What is a Measure?

We start with a simpler question: what is a “δ-function”? Typical definition:   x 6= 0  δ(x) = 0,  δ(0) = ∞    R ε δ(x)dx = 1, (for ε > 0). −ε 67

68

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

We observe: This is pure nonsense. We observe that δ(x) = 0 a.e., so Thus: The δ-function is not a function! What is it?

Rε

−ε

δ(x)dx = 0.

Normally a δ-function is used in the following way: Suppose that f is continuous at the origin. Then Z ∞ Z f (x)δ(x)dx = −∞

∞

[f (x) − f (0)] δ(x) dx + f (0) {z } |{z} −∞ |

= f (0)

Z

=0 when x=0 ∞

=0 when x6=0

Z

∞

δ(x)dx

−∞

δ(x)dx = f (0).

−∞

This gives us a new interpretation of δ: The δ-function is the “operator” which evaluates a continuous function at the point zero. Principle: You feed a function f (x) to δ, and δ gives you back the number f (0) (forget about the integral formula). Since the formal integral

R∞

−∞

f (x)δ(x)dx resembles an inner product, we often

use the notation hδ, f i. Thus

hδ, f i = f (0) Definition 3.1. The δ-operator is the (bounded linear) operator which maps f ∈ C0 (R) into the number f (0). Also called Dirac’s delta . This is a special case of measure: Definition 3.2. A measure µ is a bounded linear operator which maps functions f ∈ C0 (R) into the set of complex numbers C (or real). We denote this number by hµ, f i.

Example 3.3. The operator which maps f ∈ C0 (R) into the number Z 1 f (0) + f (1) + f (s)ds 0

is a measure. Proof. Denote hG, f i = f (0) + f (1) +

R1 0

f (s)ds. Then

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

69

i) G maps C0 (R) → C. ii) G is linear : hG, λf + µgi = λf (0) + µg(0) + λf (1) + µg(1) Z 1 + (λf (s) + µg(s))ds 0 Z 1 = λf (0) + λf (1) + λf (s)ds 0 Z 1 +µg(0) + µg(1) + µg(s)ds 0

= λhG, f i + µhG, gi.

iii) G is continuous: If fn → f in C0 (R), then maxt∈R |fn (t) − f (t)| → 0 as n → ∞, so

fn (0) → f (0), fn (1) → f (1) and

Z

0

1

fn (s)ds →

Z

1

f (s)ds,

0

so hG, fn i → hG, f i as n → ∞. Thus, G is a measure.



Warning 3.4. hG, f i is linear in f , not conjugate linear: hG, λf i = λhG, f i, and not = λhG, f i. Alternative notation 3.5. Instead of hG, f i many people write G(f ) or Gf

(for example , Gripenberg). See Gasquet for more details.

3.2

What is a Distribution?

Physisists often also use “the derivative of a δ-function”, which is defined as hδ ′ , f i = −f ′ (0), here f ′ (0) = derivative of f at zero. This is not a measure: It is not defined for all f ∈ C0 (R) (only for those that are differentiable at zero). It is linear,

but it is not continuous (easy to prove). This is an example of a more general distribution.

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

70

Definition 3.6. A tempered distribution (=tempererad distribution) is a continuous linear operator from S to C. We denote the set of such distributions

by S ′ . (The set S was defined in Section 2.2).

Theorem 3.7. Every measure is a distribution. Proof. i) Maps S into C, since S ⊂ C0 (R). ii) Linearity is OK. iii) Continuity is OK: If fn → f in S, then fn → f in C0 (R), so hµ, fn i → hµ, f i (more details below!)



Example 3.8. Define hδ ′ , ϕi = −ϕ′ (0), ϕ ∈ S. Then δ ′ is a tempered distribution

Proof. i) Maps S → C? Yes! ii) Linear? Yes! iii) Continuous? Yes! (See below for details!)



What does ϕn → ϕ in S mean? Definition 3.9. ϕn → ϕ in S means the following: For all positive integers k, m, tk ϕn(m) (t) → tk ϕ(m) (t) uniformly in t, i.e., lim max|tk (ϕn(m) (t) − ϕ(m) (t))| = 0.

n→∞ t∈R

Lemma 3.10. If ϕn → ϕ in S, then ϕn(m) → ϕ(m) in C0 (R) for all m = 0, 1, 2, . . . Proof. Obvious. Proof that δ ′ is continuous: If ϕn → ϕ in S, then maxt∈R |ϕ′n (t) − ϕ′ (t)| → 0 as n → ∞, so

hδ ′ , ϕn i = −ϕ′n (0) → ϕ′ (0) = hδ ′, ϕi. 

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

3.3

71

How to Interpret a Function as a Distribution?

Lemma 3.11. If f ∈ L1 (R) then the operator which maps ϕ ∈ S into Z ∞ hF, ϕi = f (s)ϕ(s)ds −∞

is a continuous linear map from S to C. (Thus, F is a tempered distribution). Note: No complex conjugate on ϕ! Note: F is even a measure. Proof. i) For every ϕ ∈ S, the integral converges (absolutely), and defines a number in C. Thus, F maps S → C.

ii) Linearity: for all ϕ, ψ ∈ S and λ, µ ∈ C, Z hF, λϕ + µψi = f (s)[λϕ(s) + µψ(s)]ds RZ Z = λ f (s)ϕ(s)ds + µ f (s)ψ(s)ds R

R

= λhF, ϕi + µhF, ψi.

iii) Continuity: If ϕn → ϕ in S, then ϕn → ϕ in C0 (R), and by Lebesgue’s dominated convergence theorem, Z Z hF, ϕn i = f (s)ϕn (s)ds → f (s)ϕ(s)ds = hF, ϕi.  R

R

The same proof plus a little additional work proves: Theorem 3.12. If

Z

∞

−∞

|f (t)| dt < ∞ 1 + |t|n

for some n = 0, 1, 2, . . ., then the formula Z ∞ hF, ϕi = f (s)ϕ(s)ds, −∞

defines a tempered distribution F .

ϕ ∈ S,

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

72

Definition 3.13. We call the distribution F in Lemma 3.11 and Theorem 3.12 the distribution induced by f , and often write hf, ϕi instead of hF, ϕi. Thus, Z ∞ hf, ϕi = f (s)ϕ(s)ds, ϕ ∈ S. −∞

This is sort of like an inner product, but we cannot change places of f and ϕ: f is “the distribution” and ϕ is “the test function” in hf, ϕi.

Does “the distribution f ” determine “the function f ” uniquely? Yes! Theorem 3.14. Suppose that the two functions f1 and f2 satisfy Z |fi (t)| dt < ∞ (i = 1 or i = 2), n R 1 + |t|

and that they induce the same distribution, i.e., that Z Z f1 (t)ϕ(t)dt = f2 (t)ϕ(t)dt, ϕ ∈ S. R

R

Then f1 (t) = f2 (t) almost everywhere. Proof. Let g = f1 − f2 . Then Z

R

g(t)ϕ(t)dt = 0 for all ϕ ∈ S ⇐⇒

R

g(t) (1 + t2 )n/2 ϕ(t)dt = 0 ∀ϕ ∈ S. (1 + t2 )n/2

Z

Easy to show that (1 + t2 )n/2 ϕ(t) ∈ S ⇐⇒ ϕ ∈ S. If we define h(t) = {z } |

g(t) , (1+t2 )n/2

ψ(t)

1

then h ∈ L (R), and

Z

∞

−∞

h(s)ψ(s)ds = 0 ∀ψ ∈ S.

If ψ ∈ S then also the function s 7→ ψ(t − s) belongs to S, so ( Z ∀ψ ∈ S, h(s)ψ(t − s)ds = 0 ∀t ∈ R. R

(3.1)

2

Take ψn (s) = ne−π(ns) . Then ψn ∈ S, and by 3.1, ψn ∗ h ≡ 0. On the other hand, by Theorem 2.12, ψn ∗ h → h in L1 (R) as n → ∞, so this gives h(t) = 0 a.e.



CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

73

Corollary 3.15. If we know “the distribution f ”, then from this knowledge we can reconstruct f (t) for almost all t. Proof. Use the same method as above. We know that h(t) ∈ L1 (R), and that (ψn ∗ h)(t) → h(t) =

f (t) . (1 + t2 )n/2

As soon as we know “the distribution f ”, we also know the values of Z ∞ f (s) (1 + s2 )n/2 ψn (t − s)ds (ψn ∗ h)(t) = 2 n/2 −∞ (1 + s ) for all t.

3.4



Calculus with Distributions

(=R¨akneregler)

3.16 (Addition). If f and g are two distributions, then f + g is the distribution hf + g, ϕi = hf, ϕi + hg, ϕi,

ϕ ∈ S.

(f and g distributions ⇐⇒ f ∈ S ′ and g ∈ S ′ ). 3.17 (Multiplication by a constant). If λ is a constant and f ∈ S ′ , then λf is the distribution

hλf, ϕi = λhf, ϕi,

ϕ ∈ S.

3.18 (Multiplication by a test function). If f ∈ S ′ and η ∈ S, then ηf is the

distribution

hηf, ϕi = hf, ηϕi ϕ ∈ S. Motivation: If f would be induced by a function, then this would be the natural definition, because Z Z [η(s)f (s)]ϕ(s)ds = f (s)[η(s)ϕ(s)]ds = hf, ηϕi. R

R

Warning 3.19. In general, you cannot multiply two distributions. For example, δ 2 = δδ is nonsense

(δ = “δ-function”) =Dirac’s delta

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

74

However, it is possible to multiply distributions by a larger class of “test functions”: ∞ Definition 3.20. By the class Cpol (R) of tempered test functions we mean the

following: ∞ ψ ∈ Cpol (R) ⇐⇒ f ∈ C ∞ (R),

and for every k = 0, 1, 2, . . . there are two numbers M and n so that |ψ (k) (t)| ≤ M(1 + |t|n ),

t ∈ R.

∞ Thus, f ∈ Cpol (R) ⇐⇒ f ∈ C ∞ (R), and every derivative of f grows at most as

a polynomial as t → ∞.   S = “rapidly decaying test functions”   Repetition: S ′ = “tempered distributions”    C ∞ (R) = “tempered test functions”. pol

∞ Example 3.21. Every polynomial belongs to Cpol . So do the functions

1 , 1 + x2

(1 + x2 )±m

(m need not be an integer)

∞ Lemma 3.22. If ψ ∈ Cpol (R) and ϕ ∈ S, then

ψϕ ∈ S. Proof. Easy (special case used on page 72). ∞ Definition 3.23. If ψ ∈ Cpol (R) and f ∈ S ′ , then ψf is the distribution

hψf, ϕi = hf, ψϕi,

ϕ∈S

(O.K. since ψϕ ∈ S). Now to the big surprise : Every distribution has a derivative, which is another distribution! Definition 3.24. Let f ∈ S ′ . Then the distribution derivative of f is the

distribution defined by

hf ′ , ϕi = −hf, ϕ′ i,

ϕ∈S

(This is O.K., because ϕ ∈ S =⇒ ϕ′ ∈ S, so −hf, ϕ′ i is defined).

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

75

Motivation: If f would be a function in C 1 (R) (not too big at ∞), then Z ∞ ′ hf, ϕ i = f (s)ϕ′(s)ds (integrate by parts) −∞ Z ∞ ∞ = [f (s)ϕ(s)]−∞ − f ′ (s)ϕ(s)ds | {z } −∞ =0

′

= −hf , ϕi. 

Example 3.25. Let f (t) =

(

e−t ,

t ≥ 0,

−et , t < 0.

Interpret this as a distribution, and compute its distribution derivative. Solution: ′

Z

′

∞

hf , ϕi = −hf, ϕ i = − f (s)ϕ′ (s)ds −∞ Z 0 Z ∞ s ′ = e ϕ (s)ds − e−s ϕ′ (s)ds −∞ 0 Z 0 Z ∞  −s ∞ 0 s s = [e ϕ(s)]−∞ − e ϕ(s)ds − e ϕ(s) 0 − e−s ϕ(s)ds 0 Z ∞ −∞ = 2ϕ(0) − e−|s| ϕ(s)ds. −∞

Thus, f ′ = 2δ + h, where h is the “function” h(s) = −e−|s| , s ∈ R, and δ = the Dirac delta (note that h ∈ L1 (R) ∩ C(R)).

Example 3.26. Compute the second derivative of the function in Example 3.25! Solution: By definition, hf ′′ , ϕi = −hf ′ , ϕ′ i. Put ϕ′ = ψ, and apply the rule

hf ′ , ψi = −hf, ψ ′ i. This gives

hf ′′ , ϕi = hf, ϕ′′ i. By the preceding computation ′

′

−hf, ϕ i = −2ϕ (0) −

Z

∞ −∞

e−|s| ϕ′ (s)ds

= (after an integration by parts) Z ∞ ′ = −2ϕ (0) + f (s)ϕ(s)ds −∞

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

76

(f = original function). Thus, ′′

′

hf , ϕi = −2ϕ (0) +

Z

∞

f (s)ϕ(s)ds.

−∞

Conclusion: In the distribution sense, f ′′ = 2δ ′ + f, where hδ ′ , ϕi = −ϕ′ (0). This is the distribution derivative of Dirac’s delta. In particular: f is a distribution solution of the differential equation f ′′ − f = 2δ ′ . This has something to do with the differential equation on page 59. More about this later.

3.5

The Fourier Transform of a Distribution

Repetition: By Lemma 2.19, we have Z ∞ Z f (t)ˆ g (t)dt = −∞

∞

fˆ(t)g(t)dt

−∞

if f, g ∈ L1 (R). Take g = ϕ ∈ S. Then ϕˆ ∈ S (See Theorem 2.24), so we can interpret both f and fˆ in the distribution sense and get Definition 3.27. The Fourier transform of a distribution f ∈ S is the distribu-

tion defined by

hfˆ, ϕi = hf, ϕi, ˆ

ϕ ∈ S.

Possible, since ϕ ∈ S ⇐⇒ ϕˆ ∈ S. Problem: Is this really a distribution? It is well-defined and linear, but is it continuous? To prove this we need to know that ϕn → ϕ in S ⇐⇒ ϕˆn → ϕˆ in S. This is a true statement (see Gripenberg or Gasquet for a proof), and we get Theorem 3.28. The Fourier transform maps the class of tempered distributions onto itself: f ∈ S ′ ⇐⇒ fˆ ∈ S ′ .

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

77

There is an obvious way of computing the inverse Fourier transform: Theorem 3.29. The inverse Fourier transform f of a distribution fˆ ∈ S ′ is

given by

hf, ϕi = hfˆ, ψi,

ϕ ∈ S,

where ψ = the inverse Fourier transform of ϕ, i.e., ψ(t) =

R∞

−∞

e2πitω ϕ(ω)dω.

Proof. If ψ = the inverse Fourier transform of ϕ, then ϕ = ψˆ and the formula ˆ = hf, ˆ ψi.  simply says that hf, ψi

3.6

The Fourier Transform of a Derivative

Problem 3.30. Let f ∈ S ′ . Then f ′ ∈ S ′ . Find the Fourier transform of f ′ . ∞ Solution: Define η(t) = 2πit, t ∈ R. Then η ∈ Cpol , so we can multiply a

tempered distribution by η. By various definitions (start with 3.27) d ′ ), ϕi = hf ′ , ϕi h(f ˆ

(use Definition 3.24) ′

= −hf, (ϕ) ˆ i (use Theorem 2.7(g)) ˆ = −hf, ψi (where ψ(s) = −2πisϕ(s)) = −hfˆ, ψi (by Definition 3.27)

= hfˆ, ηϕi = hη fˆ, ϕi

(see Definition above of η) (by Definition 3.23).

d ′ ) = η fˆ where η(ω) = 2πiω, ω ∈ R. Thus, (f

This proves one half of: Theorem 3.31.

d ′ ) = (i2πω)fˆ and (f \ ) = (fˆ)′ (−2πitf ∞ More precisely, if we define η(t) = 2πit, then η ∈ Cpol , and

d d) = −fˆ′ . ′ ) = η f, ˆ (ηf (f

By repeating this result several times we get Theorem 3.32.

[ (k) ) = (2πiω)k fˆ k ∈ Z (f + (k) \k f ) = fˆ . ((−2πit)

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

78

Example 3.33. Compute the Fourier transform of ( e−t , t > 0, f (t) = −et , t < 0. Smart solution: By the Examples 3.25 and 3.26. f ′′ = 2δ ′ + f

(in the distribution sense).

Transform this: d ′ ) = 2(2πiω)δ ˆ [(2πiω)2 − 1]fˆ = 2(δ

ˆ (since δ ′ is the derivative of δ). Thus, we need δ: Z ˆ hδ, ϕi = hδ, ϕi ˆ = ϕ(0) ˆ = ϕ(s)ds R Z Z = 1 · ϕ(s)ds = f (s)ϕ(s)ds, R

R

where f (s) ≡ 1. Thus δˆ is the distribution which is induced by the function f (s) ≡ 1, i.e., we may write δˆ ≡ 1 . Thus, −(4π 2 ω 2 + 1)fˆ = 4πiω, so fˆ is induced by the function fˆ(ω) =

4πiω . −(1+4π 2 ω 2 )

Thus,

4πiω . −(1 + 4π 2 ω 2 )

In particular: Lemma 3.34. ˆ δ(ω) ≡ 1 ˆ1 = δ.

and

(The Fourier transform of δ is the function ≡ 1, and the Fourier transform of the function ≡ 1 is the Dirac delta.)

Combining this with Theorem 3.32 we get Lemma 3.35. d δ (k) = (2πiω)k , h i \ k = δ (k) (−2πit)

k ∈ Z+ = 0, 1, 2, . . .

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

3.7

79

Convolutions (”Faltningar”)

It is sometimes (but not always) possible to define the convolution of two distributions. One possibility is the following: If ϕ, ψ ∈ S, then we know that \ ˆ (ϕ ∗ ψ) = ϕˆψ, ˆ The same idea so we can define ϕ ∗ ψ to be the inverse Fourier transform of ϕˆψ. applies to distributions in some cases:

Definition 3.36. Let f ∈ S ′ and suppose that g ∈ S ′ happens to be such that ∞ ∞ gˆ ∈ Cpol (R) (i.e., gˆ is induced by a function in Cpol (R), i.e., g is the inverse ∞ F -transform of a function in Cpol ). Then we define

f ∗ g = the inverse Fourier transform of fˆgˆ, i.e. (cf. page 77): hf ∗ g, ϕi = hfˆgˆ, ϕi ˇ where ϕˇ is the inverse Fourier transform of ϕ: Z ∞ ϕ(t) ˇ = e2πiωt ϕ(ω)dω. −∞

∞ This is possible since gˆ ∈ Cpol , so that fˆgˆ ∈ S ′ ; see page 74

To get a direct interpretation (which does not involve Fourier transforms) we need two more definitions: Definition 3.37. Let t ∈ R, f ∈ S ′ , ϕ ∈ S. Then the translations τt f and τt ϕ

are given by

(τt ϕ)(s) = ϕ(s − t), hτt f, ϕi = hf, τ−t ϕi

s∈R

Motivation: τt ϕ translates ϕ to the right by the amount t (if t > 0, to the left if t < 0). For ordinary functions f we have Z ∞ Z ∞ (τt f )(s)ϕ(s)ds = f (s − t)ϕ(s)ds (s − t = v) −∞ −∞ Z ∞ = f (v)ϕ(v + t)dv −∞ Z ∞ = f (v)τ−t ϕ(v)dv, −∞

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

80

ϕ τt ϕ t so the distribution definition coincides with the usual definition for functions interpreted as distributions. Definition 3.38. The reflection operator R is defined by (Rϕ)(s) = ϕ(−s), hRf, ϕi = hf, Rϕi,

ϕ ∈ S,

f ∈ S ′, ϕ ∈ S

Motivation: Extra homework. If f ∈ L1 (R) and η ∈ S, then we can write f ∗ ϕ

f 0 Rf 0 in the form (f ∗ ϕ)(t) = = =

Z

ZR

ZR

f (s)η(t − s)ds f (s)(Rη)(s − t)ds f (s)(τt Rη)(s)ds,

R

and we get an alternative formula for f ∗ η in this case. Theorem 3.39. If f ∈ S ′ and η ∈ S, then f ∗ η as defined in Definition 3.36, is induced by the function

t 7→ hf, τt Rηi, ∞ and this function belongs to Cpol (R).

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

81

We shall give a partial proof of this theorem (skipping the most complicated part). It is based on some auxiliary results which will be used later, too. Lemma 3.40. Let ϕ ∈ S, and let ϕε (t) =

ϕ(t + ε) − ϕ(t) , ε

t ∈ R.

Then ϕε → ϕ′ in S as ε → 0. Proof. (Outline) Must show that lim sup|t|k |ϕε(m) (t) − ϕ(m+1) (t)| = 0

ε→0 t∈R

for all t, m ∈ Z+ . By the mean value theorem, ϕ(m) (t + ε) = ϕ(m) (t) + εϕ(m+1) (ξ) where t < ξ < t + ε (if ε > 0). Thus |ϕε(m) (t) − ϕ(m+1) (t)| = |ϕ(m+1) (ξ) − ϕ(m+1) (t)| ! Z t where t < ξ < t + ε if ε > 0 = | ϕ(m+2) (s)ds| or t + ε < ξ < t if ε < 0 ξ Z t+|ε| ≤ |ϕ(m+2) (s)|ds, t−|ε|

and this multiplied by |t|k tends uniformly to zero as ε → 0. (Here I am skipping a couple of lines). 

Lemma 3.41. For every f ∈ S ′ there exist two numbers M > 0 and N ∈ Z+ so

that

|hf, ϕi| ≤ M max |tj ϕ(k) (t)|. 0≤j,k≤N t∈R

(3.2)

Interpretation: Every f ∈ S ′ has a finite order (we need only derivatives ϕ(k)

where k ≤ N) and a finite polynomial growth rate (we need only a finite power tj with j ≤ N).

Proof. Assume to get a contradiction that (3.2) is false. Then for all n ∈ Z+ , there is a function ϕn ∈ S so that

|hf, ϕn i| ≥ n max |tj ϕ(k) n (t)|. 0≤j,k≤n t∈R

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

82

Multiply ϕn by a constant to make hf, ϕn i = 1. Then max |tj ϕ(k) n (t)| ≤

0≤j,k≤n t∈R

1 → 0 as n → ∞, n

so ϕn → 0 in S as n → ∞. As f is continuous, this implies that hf, ϕn i → 0

as n → ∞. This contradicts the assumption hf, ϕn i = 1. Thus, (3.2) cannot be false. 

∞ Theorem 3.42. Define ϕ(t) = hf, τt Rηi. Then ϕ ∈ Cpol , and for all n ∈ Z+ ,

ϕ(n) (t) = hf (n) , τt Rηi = hf, τt Rη (n) i. Note: As soon as we have proved Theorem 3.39, we may write this as (f ∗ η)(n) = f (n) ∗ η = f ∗ η (n) . Thus, to differentiate f ∗ η it suffices to differentiate either f or η (but not both). The derivatives may also be distributed between f and η: (f ∗ η)(n) = f (k) ∗ η (n−k) ,

0 ≤ k ≤ n.

Motivation: A formal differentiation of Z (f ∗ ϕ)(t) = f (t − s)ϕ(s)ds gives R Z ′ (f ∗ ϕ) = f ′ (t − s)ϕ(s)ds = f ′ ∗ ϕ, R

and a formal differentiation of (f ∗ ϕ)(t) = (f ∗ ϕ)′ =

Z

ZR R

f (s)ϕ(t − s)ds gives f (s)ϕ′ (t − s)ds = f ∗ ϕ′ .

Proof of Theorem 3.42. i)

1 [ϕ(t ε

+ ε) − ϕ(t)] = hf, 1ε (τt+ε Rη − τt Rη)i. Here 1 [(Rη)(s − t − ε) − Rη(s − t)] ε 1 = [η(t + ε − s) − η(t − s)] (by Lemma 3.40) ε → η ′ (t − s) = (Rη ′ )(s − t) = τt Rη ′ .

1 (τt+ε Rη − τt Rη)(s) = ε

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

83

Thus, the following limit exists: 1 lim [ϕ(t + ε) − ϕ(t)] = hf, τt Rη ′ i. ε→0 ε Repeating the same argument n times we find that ϕ is n times differentiable, and that ϕ(n) = hf, τt Rη (n) i (or written differently, (f ∗ η)(n) = f ∗ η (n) .) ii) A direct computation shows: If we put ψ(s) = η(t − s) = (Rη)(s − t) = (τt Rη)(s), then ψ ′ (s) = −η ′ (t − s) = −τt Rη ′ . Thus hf, τt Rη ′ i = −hf, ψ ′ i = hf ′ , ψi =

hf ′ , τt Rηi (by the definition of distributed derivative). Thus, ϕ′ = hf, τt Rη ′ i =

hf ′ , τt Rηi (or written differently, f ∗ η ′ = f ′ ∗ η). Repeating this n times

we get

f ∗ η (n) = f (n) ∗ η. ∞ iii) The estimate which shows that ϕ ∈ Cpol : By Lemma 3.41,

|ϕ(n) (t)| = |hf (n) , τt Rηi|

≤ M max |sj (τt Rη)(k) (s)| (ψ as above) 0≤j,k≤N s∈R

= M max |sj η (k) (t − s)| (t − s = v) 0≤j,k≤N s∈R

= M max |(t − v)j η (k) (s)| 0≤j,k≤N v∈R

≤ a polynomial in |t|.  To prove Theorem 3.39 it suffices to prove the following lemma (if two distributions have the same Fourier transform, then they are equal): Lemma 3.43. Define ϕ(t) = hf, τt Rηi. Then ϕˆ = fˆηˆ. Proof. (Outline) By the distribution definition of ϕ: ˆ ˆ for all ψ ∈ S. hϕ, ˆ ψi = hϕ, ψi

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

We compute this: ˆ = hϕ, ψi =

Z

∞

−∞

Z

∞

−∞

ˆ ϕ(s) ψ(s)ds |{z}

function ∞ in Cpol

ˆ hf, τs Rηiψ(s)ds

= (this step is too difficult: To show that we may move the integral to the other side of f requires more theory then we have time to present) Z ∞ = hf, τs Rη ϕ(s)dsi ˆ = (⋆) −∞

Here τs Rη is the function (τs Rη)(t) = (Rη)(t − s) = η(s − t) = (τt η)(s), so the integral is Z ∞ Z ∞ ˆ ˆ η(s − t)ψ(s)ds = (τt η)(s)ψ(s)ds (see page 43) −∞ −∞ Z ∞ [ = (τ t η)(s)ψ(s)ds (see page 38) −∞ Z ∞ = e−2πits ηˆ(s)ψ(s)ds −∞ | {z } F -transform of ηˆψ

(⋆) = hf, ηˆc ψi = hfˆ, ηˆψi

= hfˆηˆ, ψi. Thus, ϕˆ = fˆηˆ. 

Using this result it is easy to prove: Theorem 3.44. Let f ∈ S ′ , ϕ, ψ ∈ S. Then (f ∗ ϕ) ∗ ψ = f ∗ (ϕ ∗ ψ) | {z } |{z} |{z} | {z } ∞ in S in S ′ in Cpol {z in S } {z } | | ∞ in Cpol

∞ in Cpol

Proof. Take the Fourier transforms:

(f ∗ ϕ) ∗ ψ = f ∗ (ϕ ∗ ψ) . | {z } |{z} |{z} | {z } |

↓ fˆϕ ˆ

{z

(fˆϕ) ˆ ψˆ

↓ ψˆ

}

|

↓ fˆ

{z

↓ ϕ ˆψˆ

ˆ fˆ(ϕ ˆψ)

}

84

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

85

The transforms are the same, hence so are the original distributions (note that ∞ both (f ∗ ϕ) ∗ ψ and f ∗ (ϕ ∗ ψ) are in Cpol so we are allowed to take distribution

Fourier transforms).

3.8

Convergence in S ′

We define convergence in S ′ by means of test functions in S. (This is a special case of “weak” or “weak*”-convergence).

Definition 3.45. fn → f in S ′ means that hfn , ϕi → hf, ϕi for all ϕ ∈ S. Lemma 3.46. Let η ∈ S with ηˆ(0) = 1, and define ηλ (t) = λη(λt), t ∈ R, λ > 0.

Then, for all ϕ ∈ S,

ηλ ∗ ϕ → ϕ in S as λ → ∞.

Note: We had this type of ”δ-sequences” also in the L1 -theory on page 36. Proof. (Outline.) The Fourier transform is continuous S → S (which we have

not proved, but it is true). Therefore ηλ ∗ ϕ → ϕ in S

⇐⇒

η\ ˆ in S λ ∗ϕ → ϕ

⇐⇒

ηˆλ ϕˆ → ϕˆ in S

⇐⇒

ηˆ(ω/λ)ϕ(ω) ˆ → ϕ(ω) ˆ in S as λ → ∞.

Thus, we must show that  d j sup ω k [ˆ η (ω/λ) − 1]ϕ(ω) ˆ → 0 as λ → ∞. dω ω∈R

This is a “straightforward” mechanical computation (which does take some time).  Theorem 3.47. Define ηλ as in Lemma 3.46. Then ηλ → δ in S ′ as λ → ∞. Comment: This is the reason for the name ”δ-sequence”. Proof. The claim (=”p˚ ast˚ aende”) is that for all ϕ ∈ S, Z ηλ (t)ϕ(t)dt → hδ, ϕi = ϕ(0) as λ → ∞. R

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

(Or equivalently,

R

R

86

λη(λt)ϕ(t)dt → ϕ(0) as λ → ∞). Rewrite this as Z ηλ (t)(Rϕ)(−t)dt = (ηλ ∗ Rϕ)(0), R

and by Lemma 3.46, this tends to (Rϕ)(0) = ϕ(0) as λ → ∞. Thus, hηλ , ϕi → hδ, ϕi for all ϕ ∈ S as λ → ∞, so ηλ → δ in S ′ .  Theorem 3.48. Define ηλ as in Lemma 3.46. Then, for all f ∈ S ′ , we have ηλ ∗ f → f in S ′ as λ → ∞. Proof. The claim is that hηλ ∗ f, ϕi → hf, ϕi for all ϕ ∈ S. Replace ϕ with the reflected ψ = Rϕ

=⇒

hηλ ∗ f, Rψi → hf, Rψi for all ϕ ∈ S

⇐⇒

(by Thm 3.39) ((ηλ ∗ f ) ∗ ψ)(0) → (f ∗ ψ)(0) (use Thm 3.44)

⇐⇒

f ∗ (ηλ ∗ ψ)(0) → (f ∗ ψ)(0) (use Thm 3.39)

⇐⇒

hf, R(ηλ ∗ ψ)i → hf, Rψi.

This is true because f is continuous and ηλ ∗ ψ → ψ in S, according to Lemma

3.46.

There is a General Rule about distributions: Metatheorem: All reasonable claims about distribution convergence are true. Problem: What is “reasonable”? Among others, the following results are reasonable: Theorem 3.49. All the operations on distributions and test functions which we have defined are continuous. Thus, if fn → f in S ′ , gn → g in S ′ ,

∞ ψn → ψ in Cpol (which we have not defined!),

ϕn → ϕ in S,

λn → λ in C, then, among others,

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

87

i) fn + gn → f + g in S ′ ii) λn fn → λf in S ′ iii) ψn fn → ψf in S ′ iv) ψˇn ∗ fn → ψˇ ∗ f in S ′ (ψˇ =inverse F -transform of ψ) ∞ v) ϕn ∗ fn → ϕ ∗ f in Cpol

vi) fn′ → f ′ in S ′ vii) fˆn → fˆ in S ′ etc. Proof. “Easy” but long.

3.9

Distribution Solutions of ODE:s

Example 3.50. Find the function u ∈ L2 (R+ ) ∩ C 1 (R+ ) with an “absolutely continuous” derivative u′ which satisfies the equation ( u′′ (x) − u(x) = f (x), x > 0, u(0) = 1.

Here f ∈ L2 (R+ ) is given. Solution. Let v be the solution of homework 22. Then ( v ′′ (x) − v(x) = f (x), x > 0, v(0) = 0.

Define w = u − v. Then w is a solution of ( w ′′ (x) − w(x) = 0, w(0) = 1.

x ≥ 0,

In addition we require w ∈ L2 (R+ ).

Elementary solution. The characteristic equation is λ2 − 1 = 0,

roots λ = ±1,

general solution w(x) = c1 ex + c2 e−x .

(3.3)

(3.4)

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

88

The condition w(x) ∈ L2 (R+ ) forces c1 = 0. The condition w(0) = 1 gives w(0) = c2 e0 = c2 = 1. Thus: w(x) = e−x , x ≥ 0.

Original solution: u(x) = e−x + v(x), where v is a solution of homework 22, i.e., Z Z 1 −x ∞ −y 1 ∞ −|x−y| −x u(x) = e + e e f (y)dy − e f (y)dy. 2 2 0 0 Make w an even function, and differentiate: we

Distribution solution.

denote the distribution derivatives by w (1) and w (2) . Then w (1) = w ′

(since w is continuous at zero)

w (2) = w ′′ + 2w ′ (0) δ0 | {z }

(Dirac delta at the point zero)

due to jump discontinuity at zero in w ′

The problem says: w ′′ = w, so w (2) − w = 2w ′ (0)δ0 . Transform:

((2πiγ)2 − 1)w(γ) ˆ = 2w ′ (0) (since δˆ0 ≡ 1) =⇒ w(γ) ˆ =

2w ′ (0) , 1+4π 2 γ 2

whose inverse transform is −w ′ (0)e−|x| (see page 62). We are only interested in values x ≥ 0 so

w(x) = −w ′ (0)e−x ,

x > 0.

The condition w(0) = 1 gives −w ′ (0) = 1, so w(x) = e−x ,

x ≥ 0.

Example 3.51. Solve the equation ( u′′ (x) − u(x) = f (x),

x > 0,

′

u (0) = a,

where a =given constant, f (x) given function. Many different ways exist to attack this problem: Method 1. Split u in two parts: u = v + w, where ( v ′′ (x) − v(x) = f (x), x > 0 v ′ (0) = 0,

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

and

(

w ′′(x) − w(x) = 0,

89

x>0

′

w (0) = a,

We can solve the first equation by making an even extension of v. The second equation can be solved as above. Method 2. Make an even extension of u and transform. Let u(1) and u(2) be the distribution derivatives of u. Then as above, u(1) = u′

(u is continuous)

u(2) = u′′ + 2 u′(0) δ0 | {z }

(u′ discontinuous)

=a

′′

By the equation: u = u + f , so

u(2) − u = 2aδ0 + f Transform this:

[(2πiγ)2 − 1]ˆ u = 2a + fˆ,

uˆ =

−2a 1+4π 2 γ 2

Invert:

−

so

fˆ 1+4π 2 γ 2

Z 1 ∞ −|x−y| u(x) = −ae − e f (y)dy. 2 −∞ Since f is even, this becomes for x > 0: Z Z 1 ∞ −|x−y| 1 −x ∞ −y −x e f (y)dy. e f (y)dy − u(x) = −ae − e 2 2 0 0 −|x|

Method 3. The method to make u and f even or odd works, but it is a “dirty trick” which has to be memorized. A simpler method is to define u(t) ≡ 0 and

f (t) ≡ 0 for t < 0, and to continue as above. We shall return to this method in connection with the Laplace transform.

Partial Differential Equations are solved in a similar manner. The computations become slightly more complicated, and the motivations become much more complicated. For example, we can replace all the functions in the examples on page 63 and 64 by distributions, and the results “stay the same”.

3.10

The Support and Spectrum of a Distribution

“Support” = “the piece of the real line on which the distribution stands”

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

90

Definition 3.52. The support of a continuous function ϕ is the closure (=”slutna h¨oljet”) of the set {x ∈ R : ϕ(x) 6= 0}. Note: The set {x ∈ R : ϕ(x) 6= 0} is open, but the support contains, in addition,

the boundary points of this set.

Definition 3.53. Let f ∈ S ′ and let U ⊂ R be an open set. Then f vanishes

on U (=”f¨orsvinner p˚ a U”) if hf, ϕi = 0 for all test functions ϕ ∈ S whose support is contained in U.

ϕ

U Interpretation: f has “no mass in U”, “no action on U”. Example 3.54. δ vanishes on (0, ∞) and on (−∞, 0). Likewise vanishes δ (k) (k ∈ Z+ = 0, 1, 2, . . .) on (−∞, 0) ∪ (0, ∞). Proof. Obvious. Example 3.55. The function f (t) =

(

1 − |t|, |t| ≤ 1, 0, |t| > 1,

vanishes on (−∞, −1) and on (1, ∞). The support of this function is [−1, 1]

(note that the end points are included ).

Definition 3.56. Let f ∈ S ′ . Then the support of f is the complement of the

largest set on which f vanishes. Thus,  M is closed, f vanishes on R \ M, and    supp(f ) = M ⇔ f does not vanish on any open set Ω    which is strictly bigger than R \ M. (k)

Example 3.57. The support of the distribution δa is the single point {a}. Here

k ∈ Z+ , and δa is point evaluation at a:

hδa , ϕi = ϕ(a).

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

91

ˆ Definition 3.58. The spectrum of a distribution f ∈ S ′ is the support of f. Lemma 3.59. If M ⊂ R is closed, then supp(f ) ⊂ M if and only if f vanishes on R \ M.

Proof. Easy. 1 (n) , as Example 3.60. Interpret f (t) = tn as a distribution. Then fˆ = (−2πi) nδ we saw on page 78. Thus the support of fˆ is {0}, so the spectrum of f is {0}.

By adding such functions we get: Theorem 3.61. The spectrum of the function f (t) ≡ 0 is empty. The spectrum of every other polynomial is the single point {0}.

Proof. f (t) ≡ 0 ⇐⇒ spectrum is empty follows from definition. The other half is proved above. 

The converse is true, but much harder to prove: Theorem 3.62. If f ∈ S ′ and if the spectrum of f is {0}, then f is a polynomial

(6≡ 0).

This follows from the following theorem by taking Fourier transforms: Theorem 3.63. If the support of f is one single point {a} then f can be written as a finite sum

f=

n X

an δa(k) .

k=0

Proof. Too difficult to include. See e.g., Rudin’s “Functional Analysis”. Possible homework: Show that Theorem 3.64. The spectrum of f is {a} ⇐⇒ f (t) = e2πiat P (t), where P is a polynomial, P 6≡ 0.

Theorem 3.65. Suppose that f ∈ S ′ has a bounded support, i.e., f vanishes on (−∞, −T ) and on (T, ∞) for some T > 0 ( ⇐⇒ supp(f ) ⊂ [−T, T ]). Then fˆ

can be interpreted as a function, namely as

ˆ f(ω) = hf, η(t)e−2πiωt i, where η ∈ S is an arbitrary function satifying η(t) ≡ 1 for t ∈ [−T −1, T +1] (or,

more generally, for t ∈ U where U is an open set containing supp(f )). Moreover, fˆ ∈ C ∞ (R). pol

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

92

Proof. (Not quite complete) Step 1. Define ψ(ω) = hf, η(t)e−2πiωt i, where η is as above. If we choose two different η1 and η2 , then η1 (t) − η2 (t) = 0 is an open set U containing supp(f ). Since f vanishes on R\U, we have hf, η1 (t)e−2πiωt i = hf, η2 (t)e−2πiωt i, so ψ(ω) does not depend on how we choose η. Step 2. For simplicity, choose η(t) so that η(t) ≡ 0 for |t| > T + 1 (where T as

in the theorem). A “simple” but boring computation shows that

1 −2πi(ω+ε)t ∂ −2πiωt [e − e−2πiωt ]η(t) → e η(t) = −2πite−2πiωt η(t) ε ∂ω in S as ε → 0 (all derivatives converge uniformly on [−T −1, T +1], and everything

is ≡ 0 outside this interval). Since we have convergence in S, also the following

limit exists:

1 lim (ψ(ω + ε) − ψ(ω)) = ψ ′ (ω) ε→0 ε

1 = limhf, (e−2πi(ω+ε)t − e−2πiωt )η(t)i ε→0 ε = hf, −2πite−2πiωt η(t)i.

Repeating the same computation with η replaced by (−2πit)η(t), etc., we find that ψ is infinitely many times differentiable, and that ψ (k) (ω) = hf, (−2πit)k e−2πiωt η(t)i,

k ∈ Z+ .

(3.5)

Step 3. Show that the derivatives grow at most polynomially. By Lemma 3.41, we have |hf, ϕi| ≤ M max |tj ϕ(l) (t)|. 0≤t,l≤N t∈R

Apply this to (3.5) =⇒ |ψ

(k)

 d l (−2πit)k e−2πiωt η(t) . (ω)| ≤ M max tj 0≤j,l≤N dt t∈R

The derivative l = 0 gives a constant independent of ω. The derivative l = 1 gives a constant times |ω|.

93

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

The derivative l = 2 gives a constant times |ω|2, etc.

∞ Thus, |ψ (k) (ω)| ≤ const. x[1 + |w|N ], so ψ ∈ Cpol . Step 4. Show that ψ = fˆ. That is, show that Z ψ(ω)ϕ(ω)dω = hfˆ, ϕi(= hf, ϕi). ˆ R

Here we need the same “advanced” step as on page 83: Z Z ψ(ω)ϕ(ω)dω = hf, e−2πiωt η(t)ϕ(ω)idω R R Z = (why??) = hf, e−2πiωt η(t)ϕ(ω)dωi R

= hf, η(t)ϕ(t)i ˆ

since η(t) ≡ 1 in a

neighborhood of supp(f )

!

= hf, ϕi. ˆ A very short explanation of why “why??” is permitted: Replace the integral by a Riemann sum, which converges in S, i.e., approximate Z ∞ X 1 e−2πiωk t ϕ(ωk ) , e−2πiωt ϕ(ω)dω = lim n→∞ n R k=−∞

where ωk = k/n.

3.11

Trigonometric Polynomials

Definition 3.66. A trigonometric polynomial is a sum of the form ψ(t) =

m X

cj e2πiωj t .

j=1

The numbers ωj are called the frequencies of ψ. Theorem 3.67. If we interpret ψ as a polynomial then the spectrum of ψ is {ω1 , ω2 , . . . , ωm }, i.e., the spectrum consists of the frequencies of the polynomial. Proof. Follows from homework 27, since supp(δωj ) = {ωj }.  Example 3.68. Find the spectrum of the Weierstrass function σ(t) =

∞ X k=0

where 0 < a < 1, ab ≥ 1.

ak cos(2πbk t),

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

94

To solve this we need the following lemma Lemma 3.69. Let 0 < a < 1, b > 0. Then N X k=0

k

k

a cos(2πb t) →

∞ X

ak cos(2πbk t)

k=0

in S ′ as N → ∞. Proof. Easy. Must show that for all ϕ ∈ S, ! Z X N ∞ X ak cos(2πbk t)ϕ(t)dt → 0 as N → ∞. − R

k=0

k=0

This is true because Z X Z ∞ k k |a cos(2πb t)ϕ(t)|dt ≤ R k=N +1

≤ = PN

P∞

∞ X

ak |ϕ(t)|dt

R k=N +1 Z ∞ ∞ X k

a

k=N +1 N +1 Z ∞

a 1−a

−∞

−∞

|ϕ(t)|dt

|ϕ(t)|dt → 0 as N → ∞.

in S ′ , also the Fourier transforms converge P k k in S ′ , so to find σ ˆ it is enough to find the transform of N k=0 a cos(2πb t) and

Solution of 3.68: Since

k=0

→

k=0

to let N → ∞. This transform is

1 δ0 + [a(δb + δ−b ) + a2 (δ−b2 + δb2 ) + . . . + aN (δ−bN + δbN )]. 2 Thus, σˆ = δ0 +

∞

1X n a (δ−bn + δbn ), 2 n=1

where the sum converges in S ′ , and the support of this is {0, ±b, ±b2 , ±b3 , . . .}, which is also the spectrum of σ.

Example 3.70. Let f be periodic with period 1, and suppose that f ∈ L1 (T), R1 i.e., 0 |f (t)|dt < ∞. Find the Fourier transform and the spectrum of f . Solution: (Outline) The inversion formula for the periodic transform says that f=

∞ X

n=−∞

fˆ(n)e2πint .

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

95

Working as on page 86 (but a little bit harder) we find that the sum converges in S ′ , so we are allowed to take transforms: fˆ =

∞ X

ˆ f(n)δ n

n=−∞

(converges still in S ′ ).

ˆ Thus, the spectrum of f is {n ∈ N : f(n) 6= 0}. Compare this to the theory of Fourier series.

General Conclusion 3.71. The distribution Fourier transform contains all the other Fourier transforms in this course. A “universal transform”.

3.12

Singular differential equations

Definition 3.72. A linear differential equation of the type n X

ak u(k) = f

(3.6)

k=0

is regular if it has exactly one solution u ∈ S ′ for every f ∈ S ′ . Otherwise it is singular.

Thus: Singular means: For some f ∈ S ′ it has either no solution or more than

one solution.

Example 3.73. The equation u′ = f . Taking f = 0 we get many different solutions, namely u =constant (different constants give different solutions). Thus, this equation is singular. Example 3.74. We saw earlier on page 59-63 that if we work with L2 -functions instead of distributions, then the equation u′′ + λu = f is singular iff λ > 0. The same result is true for distributions: Theorem 3.75. The equation (3.6) is regular ⇔

n X k=0

ak (2πiω)k 6= 0

for all ω ∈ R.

96

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

Before proving this, let us define Definition 3.76. The function D(ω) = (3.6)

Pn

k=0

ak (2πiω)k is called the symbol of

Thus: Singular ⇔ the symbol vanishes for some ω ∈ R.

Proof of theorem 3.75. Part 1: Suppose that D(ω) 6= 0 for all ω ∈ R.

Transform (3.6):

n X k=0

ak (2πiω)k uˆ = fˆ ⇔

If D(ω) 6= 0 for all ω, then

1 D(ω)

uˆ =

D(ω)ˆ u = fˆ.

∞ ∈ Cpol , so we can multiply by

1 ˆ f D(ω)

⇔

1 : D(ω)

u=K ∗f

where K is the inverse distribution Fourier transform of

1 . D(ω)

Therefore, (3.6)

has exactly one solution u ∈ S ′ for every f ∈ S ′ .

Part 2: Suppose that D(a) = 0 for some a ∈ R. Then hDδa , ϕi = hδa , Dϕi = D(a)ϕ(a) = 0. This is true for all ϕ ∈ S, so Dδa is the zero distribution: Dδa = 0. ⇔

n X

ak (2πiω)k δa = 0.

k=0

Let v be the inverse transform of δa , i.e., 2πiat

v(t) = e

⇒

n X

ak v (k) = 0.

k=0

Thus, v is one solution of (3.6) with f ≡ 0. Another solution is v ≡ 0. Thus,

(3.6) has at least two different solutions ⇒ the equation is singular.



Definition 3.77. If (3.6) is regular, then we call K =inverse transform of

1 D(ω)

the Green’s function of (3.6). (Not defined for singular problems.) How many solutions does a singular equation have? Find them all! (Solution later!)

97

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

Example 3.78. If f ∈ C(R) (and |f (t)| ≤ M(1 + |t|k ) for some M and k), then

the equation

u′ = f has at least the solutions u(x) =

Z

x

f (x)dx + constant 0

Does it have more solutions? Answer: No! Why? Suppose that u′ = f and v ′ = f ⇒ u′ −v ′ = 0. Transform this ⇒ (2πiω)(ˆ u −ˆ v) = 0.

Let ϕ be a test function which vanishes in some interval [−ε, ε](⇔ the support of ϕ is in (−∞, −ε] ∪ [ε, ∞)). Then ψ(ω) =

ϕ(ω) 2πiω

is also a test function (it is ≡ 0 in [−ε, ε]), since (2πiω)(ˆ u − vˆ) = 0 we get 0 = h(2πiω)(ˆ u − vˆ), ψi = hˆ u − vˆ, 2πiωψ(ω)i = hˆ u − vˆ, ϕi. Thus, hˆ u − vˆ, ϕi = 0 when supp(ϕ) ⊂ (−∞, −ε] ∪ [ε, ∞), so by definition,

supp(ˆ u − vˆ) ⊂ {0}. By theorem 3.63, uˆ − vˆ is a polynomial. The only polynomial

whose derivative is zero is the constant function, so u − v is a constant.



A more sophisticated version of the same argument proves the following theorem: Theorem 3.79. Suppose that the equation n X

ak u(k) = f

(3.7)

k=0

is singular, and suppose that the symbol D(ω) has exactly r simple zeros ω1 , ω2 , . . . , ωr . If the equation (3.7) has a solution v, then every other solution u ∈ S ′ of (3.7) is of the form

u=v+

r X

bj e2πiωj t ,

j=1

where the coefficients bj can be chosen arbitrarily.

CHAPTER 3. FOURIER TRANSFORMS OF DISTRIBUTIONS

98

Compare this to example 3.78: The symbol of the equation u′ = f is 2πiω, which has a simple zero at zero. Comment 3.80. The equation (3.7) always has a distribution solution, for all f ∈ S ′ . This is proved for the equation u′ = f in [GW99, p. 277], and this can be extended to the general case.

Comment 3.81. A zero ωj of order r ≥ 2 of D gives rise to terms of the type P (t)e2πiωj t , where P (t) is a polynomial of degree ≤ r − 1.

Chapter 4 The Fourier Transform of a Sequence (Discrete Time) From our earlier results we very quickly get a Fourier transform theory for sequences {an }∞ n=−∞ . We interpret this sequence as the distribution ∞ X

an δn

(δn = Dirac’s delta at the point n)

n=−∞

For example, this converges in S ′ if |an | ≤ M(1 + |n|N )

for some M, N

and the Fourier transform is: ∞ X

−2πiωn

an e

=

n=−∞

∞ X

k=−∞

a−k e2πiωk

which also converges in S ′ . This transform is identical to the inverse transform

discussed in Chapter 1 (periodic function!), except for the fact that we replace i by −i (or equivalently, replace n by −n). Therefore:

Theorem 4.1. All the results listed in Chapter 1 can be applied to the theory of Fourier transforms of sequences, provided that we intercharge the Fourier transform and the inverse Fourier transform. Notation 4.2. To simplify the notations we write the original sequence as f (n), n ∈ Z, and denote the Fourier transform as fˆ. Then fˆ is periodic (function or 99

100

CHAPTER 4. FOURIER SERIES

distribution, depending on the size of |f (n)| as n → ∞), and fˆ(ω) =

∞ X

f (n)e−2πiωn .

n=−∞

From Chapter 1 we can give e.g., the following results: Theorem 4.3. i) f ∈ ℓ2 (Z) ⇔ fˆ ∈ L2 (T), ii) f ∈ ℓ1 (Z) ⇒ fˆ ∈ C(T) (converse false), ( ( ˆ ∈ L1 (T) f f ∈ ℓ2 (Z) or iii) (fcg) = fˆ ∗ gˆ if e.g. gˆ ∈ L1 (T) g ∈ ℓ2 (Z) iv) Etc.

We can also define discrete convolutions: P Definition 4.4. (f ∗ g)(n) = ∞ k=−∞ f (n − k)g(k).

This is defined whenever the sum converges absolutely. For example, if f (k) 6= 0 only for finitely many k or if

f ∈ ℓ1 (Z), g ∈ ℓ∞ (Z), or if f ∈ ℓ2 (Z), g ∈ ℓ2 (Z), etc.

Lemma 4.5. i) f ∈ ℓ1 (Z), g ∈ Lp (Z), 1 ≤ p ≤ ∞, ⇒ f ∗ g ∈ ℓp (Z) ii) f ∈ ℓ1 (Z), g ∈ c0 (Z) ⇒ f ∗ g ∈ c0 (Z). Proof. “Same” as in Chapter 1 (replace all integrals by sums). Theorem 4.6. If f ∈ ℓ1 (Z) and g ∈ ℓ1 (Z), then (f[ ∗ g)(ω) = fˆ(ω)ˆ g(ω). Also true if e.g. f ∈ ℓ2 (Z) and g ∈ ℓ2 (Z). Proof. ℓ1 -case: “Same” as proof of Theorem 1.21 (replace integrals by sums). In the ℓ2 -case we first approximate by an ℓ1 -sequence, use the ℓ1 -theory, and pass to the limit.

101

CHAPTER 4. FOURIER SERIES

Notation 4.7. Especially in the engineering literature, but also in mathematical literature, one often makes a change of variable: we have fˆ(ω) = =

∞ X

n=−∞ ∞ X

−2πiωn

f (n)e

∞ X

=

f (n) e−2πiω

n=−∞


n

f (n)z −n ,

n=−∞

where z = e2πiω . Definition 4.8. Engineers define F (z) = (=“dubbelsidig”) Z-transformation of f .

P∞

n=−∞

f (n)z −n as the (bilateral)

Definition 4.9. Most mathematicians define F (z) =

P∞

n=−∞

f (n)z n instead.

Note: If f (n) = 0 for n < 0 we get the onesided (=unilateral) transform F (z) =

∞ X n=0

f (n)z

−n

(or

∞ X

f (n)z n ).

n=0

Note: The Z-transform is reduced to the Fourier transform by a change of variable z = e2πiω , so ω ∈ [0, 1] ⇔ |z| = 1 Thus, z takes values on the unit circle. In the case of one-sided sequences we can also allow |z| > 1 (engineers) or |z| < 1 (mathematicians) and get power series

like those studied in the theory of analytic functions.

All Fourier transform results apply

Chapter 5 The Discrete Fourier Transform We have studied four types of Fourier transforms: i) Periodic functions on R ⇒ fˆ defined on Z. ii) Non-periodic functions on R ⇒ fˆ defined on R. iii) Distributions on R ⇒ fˆ defined on R. iv) Sequences defined on Z ⇒ fˆ periodic on R. The final addition comes now: v) f a periodic sequence (on Z) ⇒ fˆ a periodic sequence.

5.1

Definitions

Definition 5.1. ΠN = {all periodic sequences F (m) with period N, i.e., F (m + N) = F (m)}.

Note: These are in principle defined for all n ∈ Z, but the periodicity means

that it is enough to know F (0), F (1), . . . , F (N − 1) to know the whole sequence (or any other set of N consecutive (= p˚ a varandra f¨oljande) values).

Definition 5.2. The Fourier transform of a sequence F ∈ ΠN is given by N −1 1 X − 2πimk ˆ e N F (k), m ∈ Z. F (m) = N k=0

102

103

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

1 N

Warning 5.3. Some people replace the constant

in front of the sum by

√1 N

or

omit it completely. (This affects the inversion formula.) Lemma 5.4. Fˆ is periodic with the same period N as F . Proof. Fˆ (m + N) =

1 N

1 = N

X

e−

2πi(m+N)k N

F (k)

one period

X

one period

= Fˆ (m). 

−2πik − |e {z } e

2πimk N

F (k)

=1

Thus, F ∈ ΠN ⇒ Fˆ ∈ ΠN . Theorem 5.5. F can be reconstructed from Fˆ by the inversion formula F (k) =

N −1 X

e

2πimk N

Fˆ (m).

m=0

Note: No

1 N

in front here.

Note: Matlab puts the

1 N

in front of the inversion formula instead!

Proof. X m

e

2πimk N

N −1 1 X − 2πiml 1 X e N F (l) = F (l) N l N l=0

N −1 X

=

e

m=0 | 8 > > < N, > > :

0,

2πim(k−l) N

{z

= F (k)

}

if l = k if l 6= k

 2πi N 2πi = 1, so e N is the N:th root of 1: We know that e N e2*2πi/N 2π i/N e 2π/Ν

2π/Ν 1

We add N numbers, whose absolute value is one, and who point symmetrically in all the different directions indicated above. For symmetry reasons, the sum

104

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

must be zero (except when l = k). (You always jump an angle

2π(k−l) N

for each

turn, and go k − l times around before you are done.) Definition 5.6. The convolution F ∗ G of two sequences in ΠN are defined by X F (m − k)G(k) (F ∗ G)(m) = one period

(Note: Some indeces get out of the interval [0, N −1]. You must use the periodicity of F and G to get the corresponding values of F (m − k)G(k).). Definition 5.7. The (ordinary) product F · G is defined by (F · G)(m) = F (m)G(m), m ∈ Z. ˆ and (F ˆ (note the extra factor \ Theorem 5.8. (F[ · G) = Fˆ ∗ G ∗ G) = N Fˆ · G N).

Proof. Easy. (Homework?) Definition 5.9. (RF )(n) = F (−n) (reflection operator). As before: The inverse transform = the ususal transform plus reflection: d ) (note the extra factor N), where ˆ = Fourier Theorem 5.10. Fˇ = N(RF

transform and ˇ= Inverse Fourier transform.

Proof. Easy. We could have avoided the factor N by a different scaling (but then it shows up in other places instead).

5.2

FFT=the Fast Fourier Transform

Question 5.11. How many flops do we need to compute the Fourier transform of F ∈ ΠN ? FLOP=FLoating Point Operation={multiplication or addition or combination of both}. 1 Megaflop = 1 million flops/second (106 ) 1 Gigaflop = 1 billion flops/second (109) (Used as speed measures of computers.)

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

Task 5.12. Compute Fˆ (m) = of flops (=quickly).

1 N

PN −1 k=0

e−

2πimk N

105

F (k) with the minimum amount

 2πi k Good Idea 5.13. Compute the coefficients e− N = ω k only once, and store them in a table. Since ω k+N = ω k , we have e−

2πimk N

= ω mk = ω r where r =

remainder when we divide mk by N. Thus, only N numbers need to be stored. Thus: We can ignore the number of flops needed to compute the coefficients e−

2πimk N

(done in advance).

Trivial Solution 5.14. If we count multiplication and addition separetely, then we need to compute N coefficients (as m = 0, 1, . . . , N − 1), and each coefficient

requires N muliplications and N − 1 additions. This totals

N(2N − 1) = 2N 2 − N ≈ 2N 2 flops . This is too much. Brilliant Idea 5.15. Regroup (=omgruppera) the terms, using the symmetry. Start by doing even coefficients and odd coefficients separetely: Suppose for simplicity that N is even. Then, for even m, (put N = 2n) Fˆ (2m) =

N −1 1 X 2mk ω F (k) N k=0

n−1 2n−1 i X 1 hX 2mk 2mk ω F (k) + ω F (k) = N k=0 k=n | {z } Replace k by k+n

# " n−1 1 X 2mk ω F (k) + ω 2m(k+n) F (k + n) = N k=0 =

n−1 2 X − 2πimk 1 e (N/2) [F (k) + F (k + n)] . N k=0 2

This is a new discrete time periodic Fourier transform of the sequence G(k) = N 1 [F (k) + F (n + k)] with period n = . 2 2 A similar computation (see Gripenberg) shows that the odd coefficients can be computed from n−1

1 X − 2πimk e n H(k), Fˆ (2m + 1) = n k=0

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

where H(k) = 21 e−

iπk n

106

[F (k) − F (k + n)]. Thus, instead of one transform of order

N we get two transforms of order n =

N . 2

Number of flops: Computing the new transforms by brute force (as in 5.14 on page 105) we need the following flops: Even: n(2n − 1) =

N2 − N2 2 − iπk n

Odd: The numbers e

+ n additions = − 2iπk N

=e

N2 2

flops.

are found in the table already computed.

We essentially again need the same amount, namely

N2 2

+ N2 (n extra multiplica-

tions). Total:

N2 2

2

+ N2 + N2 = N 2 + N2 ≈ N 2 . Thus, this approximately halfed the number

of needed flops. Repeat 5.16. Divide the new smaller transforms into two halfs, and again, and again. This is possible if N = 2k for some integer k, e.g., N = 1024 = 210 . Final conclusion: After some smaller adjustments we get down to 3 k 2 k flops. 2 Here N = 2k , so k = log2 N, and we get Theorem 5.17. The Fast Fourier Transform with radius 2 outlined above needs approximately 23 N log2 N flops. This is much smaller than 2N 2 − N for large N. For example N = 210 = 1024 gives

3 N log2 N ≈ 15000 ≪ 2000000 = 2N 2 − N. 2

Definition 5.18. Fast Fourier transform with  k    radius 2 : split into 2 parts at each step N = 2 radius 3 : split into 3 parts at each step N = 3k    radius m : split into m parts at each step N = mk

Note: Based on symmetries. “The same” computations repeat themselves, so by combining them in a clever way we can do it quicker. Note: The FFT is so fast that it caused a minor revolution to many branches of

numerical analysis. It made it possible to compute Fourier transforms in practice. Rest of this chapter: How to use the FFT to compute the other transforms discussed earlier.

107

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

5.3

Computation of the Fourier Coefficients of a Periodic Function

Problem 5.19. Let f ∈ C(T). Compute Z 1 ˆ f (k) = e−2πikt f (t)dt 0

as efficiently as possible. Solution: Turn f into a periodic sequence and use FFT! Conversion 5.20. Choose some N ∈ Z, and put F (m) = f (

m ), m ∈ Z N

(equidistant “sampling”). The periodicity of f makes F periodic with period N. Thus, F ∈ ΠN . Theorem 5.21 (Error estimte). If f ∈ C(T) and fˆ ∈ ℓ1 (Z) (i.e.,

then

Fˆ (m) − fˆ(m) =

X

ˆ + kN). f(m

P ˆ |f(k)| < ∞),

k6=0

Proof. By the inversion formula, for all t, X ˆ f (t) = e2πijt f(j). j∈Z

Put tk =

k N

⇒

f (tk ) = F (k) =

X

e

2πikj N

fˆ(j)

j∈Z

(this series converges uniformly by Lemma 1.14). By the definition of Fˆ : Fˆ (m) =

N −1 1 X − 2πimk e N F (k) N k=0

N −1 X

1 Xˆ f(j) = N j∈Z

=

=

X l∈Z

8 > > < > > :

ˆ + Nl). f(m

2πi(j−m)k N

|k=0 {z

N, if 0,

e

if

j−m N j−m N

}

= integer 6= integer

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

108

Take away the term fˆ(m) (l = 0) to get ˆ Fˆ (m) = f(m) +

X

fˆ(m + Nl).

l6=0

ˆ Note: If N is “large” and if f(m) → 0 “quickly” as m → ∞, then the error X l6=0

fˆ(m + Nl) ≈ 0.

First Method 5.22. Put i) fˆ(m) ≈ Fˆ (m) if |m| <

N 2

ii) fˆ(m) ≈ 12 Fˆ (m) if |m| = iii) fˆ(m) ≈ 0 if |m| >

N 2

(N even)

N . 2

ˆ N ) = Fˆ (m) instead. Here ii) is not important. We could use fˆ( N2 ) = 0 or f( 2 Here

N −1 1 X − 2πimk ˆ e N F (k). F (m) = N k=0

Notation 5.23. Let us denote (note the extra star) ∗ X

ak =

|k|≤N/2

X

ak

|k|≤N/2

= the usual sum of ak if N odd (then we have exactly N terms), and a sum where the first and last terms have been ∗ X

|k|≤N/2

ak =

divided by two (these are the same if the sequence is periodic with period N, there is “one term too many” in this case).

First Method 5.24 (Error). The first method gives the error: i) |m| <

N 2

gives the error ˆ |f(m) − Fˆ (m)| ≤

ii) |m| =

N 2

X k6=0

ˆ + kN)| |f(m

gives the error 1 ˆ |f(m) − Fˆ (m)| 2

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

iii) |m| >

N 2

109

ˆ gives the error |f(m)|.

we can simplify this into the following crude (=”grov”) estimate: sup |fˆ(m) − Fˆ (m)| ≤

m∈Z

∗ X

|fˆ(m)|

(5.1)

|m|≥N/2

(because this sum is ≥ the actual error). First Method 5.25 (Drawbacks). 1◦ Large error. 2◦ Inaccurate error estimate (5.1). 3◦ The error estimate based on fˆ and not on f . We need a better method. Second Method 5.26 (General Principle). 1◦ Evaluate t at the points tk =

k N

(as before), F (k) = f (tk )

2◦ Use the sequence F to construct a new function P ∈ C(T ) which “approximates” f .

3◦ Compute the Fourier coefficients of P . ˆ 4◦ Approximate f(n) by Pˆ (n). For this to succeed we must choose P in a smart way. The final result will be quite simple, but for later use we shall derive P from some “basic principles”. Choice of P 5.27. Clearly P depends on F . To simplify the computations we require P to satisfy (write P = P (F )) A) P is linear: P (λF + µG) = λP (F ) + µP (G) B) P is translation invariant: If we translate F , then P (F ) is translated by the same amount: If we denote (τj F )(m) = F (m − j), then P (τj F ) = τj/N P (F ) (j discrete steps ⇐⇒ a time difference of j/N).

110

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

This leads to simple computations: We want to compute Pˆ (m) (which we use as approximations of fˆ(m)) Define a δ-sequence: D(n) =

(

1, for n = 0, ±N, ±2N, . . . 0, otherwise.

Then

zero

N

period

(τk D)(n) =

(

2N

1, if n = k + jN, j ∈ Z 0, otherwise,

so [F (k)τk D](n) =

(

F (k), n = k + jN 0,

otherwise.

and so F =

N −1 X

F (k)τk D

k=0

Therefore, the principles A) and B) give P (F ) = =

N −1 X

k=0 N −1 X

F (k)P (τk D) F (k)τk/N P (D),

k=0

Where P (D) is the approximation of D = “unit pulse at time zero” D.

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

111

We denote this function by p. Let us transform P (F ): (P[ (F ))(m) =

Z

0

= = =

+1 1N X

N +1 X

k=0 N +1 X

k=0 N +1 X

F (k)(τk/N p)(s)e−2πism ds

k=0

F (k)

Z

0

F (k)

1

e−2πism p(s −

Z

k

one period

e−2πim(t+ N ) p(t)dt

− −2πimk N

F (k)e

k=0

= pˆ(m)

N +1 X

|k=0

k k )ds (s − = t) N N

F (k)e− {z

=N Fˆ (m)

= N pˆ(m)Fˆ (m).

Z |

one period

e−2πimt p(t)dt {z

pˆ(m)

2πimk N

}

}

We can get rid of the factor N by replacing p by Np. This is our approximation of the “pulse of size N at zero” (

N, n = 0 + jN 0,

otherwise.

Second Method 5.28. Construct F as in the First Method, and compute Fˆ . Then the approximation of fˆ(m) is fˆ(m) ≈ Fˆ (m)ˆ p(m), where pˆ is the Fourier transform of the function that we get when we apply our approximation procedure to the sequence ( N, n = 0(+jN) ND(n) = 0, otherwise. Note: The complicated proof of this simple method will pay off in a second! Approximation Method 5.29. Use any type of translation-invariant interpolation method, for example splines. The simplest possible method is linear interpolation: If we interpolate the pulse ND in this way we get Thus,

112

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

N

N points

p(t) =

(

N(1 − N|t|), |t| ≤ 1 N

0,

(periodic extension)

1 N

≤ |t| ≤ 1 −

1 N

This is a periodic version of the kernel. A direct computation gives  2 sin(πm/N) pˆ(m) = . πm/N We get the following interesting theorem: Theorem 5.30. If we first discretize f , i.e. we replace f by the sequence F (k) = f (k/N), the compute Fˆ (m), and finally multiply Fˆ (m) by pˆ(m) =



sin(πm/N) πm/N

2

.

then we get the Fourier coefficients for the function which we get from f by linear interpolation at the points tk = k/N. (This corresponds to the computation of the Fourier integral

R1 0

e−2πimt f (t)dt by

using the trapetsoidal rule. Other integration methods have similar interpretations.)

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

5.4

113

Trigonometric Interpolation

Problem 5.31. Construct a good method to approximate a periodic function f ∈ C(T ) by a trigonometric polynomial N X

am e2πimt

m=−N

(a finite sum, resembles inverse Fourier transformation). Useful for numerical computation etc. Note: The earlier “Second Method” gave us a linear interpolation, not trigonometric approximation. Note: This trigonometric polynomial has only finitely many Fourier coefficients 6= 0 (namely am , |m| ≤ N).

Actually, the “First Method” gave us a trigonometric polynomial. There we had  N ˆ ˆ    f (m) ≈ F (m) for |m| < 2 , fˆ(m) ≈ 21 Fˆ (m) for |m| = N2 ,    fˆ(m) ≈ 0 for |m| > N . 2

By inverting this sequence we get a trigonometric approximation of f : f (t) ≈ g(t), where g(t) =

∗ X

Fˆ (m)e2πimt .

(5.2)

|m|≤N/2

We have two different errors: i) fˆ(m) is replaced by Fˆ (m) =

1 N

PN −1 k=0

f ( Nk )e

2πikm N

,

ii) The inverse series was truncated to N terms. Strange fact: These two errors (partially) cancel each other. Theorem 5.32. The function g defined in (5.2) satisfies g(

k k ) = f ( ), n ∈ Z, N N

i.e., g interpolates f at the points tk (which were used to construct first F and then g).

114

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

Proof. We defined F (k) = f ( Nk ), and ∗ X

Fˆ (m) =

F (k)e−

2πimk N

.

|k|≤N/2

By the inversion formula on page 103, ∗ X

k g( ) = N

2πimk Fˆ (m)e N (use periodicity)

|m|≤N/2 N −1 X

=

2πimk Fˆ (m)e N

m=0

= F (k) = f (

k ) N



Error estimate: How large is |f (t)−g(t)| between the mesh points tk =

k N

(where

the error is zero)? We get an estimate from the computation in the last section. Suppose that fˆ ∈ ℓ1 (Z) and f ∈ C(T ) so that the inversion formula holds for all

t (see Theorem 1.37). Then

f (t) =

X

fˆ(m)e2πimt , and

m∈Z

∗ X

g(t) =

Fˆ (m)e2πimt (Theorem 5.21)

|m|≤N/2

=

∗ X

"

fˆ(m) +

|m|≤N/2

= f (t) −

X

#

fˆ(m + kN) e2πimt

k6=0

∗ X

∗ X X

2πimt ˆ f(m)e +

ˆ + kN)e2πimt . f(m

|m|≤N/2 k6=0

|m|≥N/2

Thus |g(t) − f (t)| ≤

∗ X

|m|≥N/2

ˆ |f(m)| +

(take l = m + kN, every |l| > |l| =

N 2

two times).

N 2

∗ X X

|m|≤N/2 k6=0

|

=

P∗

ˆ + kM)| = 2 |f(m {z

ˆ

|l|≥N/2 |f (l)|

}

appears one time, no |l| <

This leads to the following theorem:

∗ X

ˆ |f(m)|

|m|≥N/2

N 2

appears, and

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

Theorem 5.33. If

P∞

ˆ

m=−∞ |f(m)|

< ∞, then

|g(t) − f (t)| ≤ 2 where

P∗

g(t) = Fˆ (m) =

1 N

115

|m|≤N/2

P∗

∗ X

|fˆ(m)|,

|m|≥N/2

Fˆ (m)e2πimt , and

|m|≤N/2

e−

2πimk N

f ( Nk ).

This is nice if fˆ(m) → 0 rapidly as m → ∞. Better accuracy by increasing N.

5.5

Generating Functions

Definition 5.34. The generating function of the sequence Jn (x) is the function f (x, z) =

X

Jn (x)z n ,

n

where the sum over n ∈ Z or over n ∈ Z+ , depending on for which values of n the functions Jn (x) are defined.

Note: We did this in the course on special functions. E.g., if Jn = Bessel’s function of order n, then x

f (x, z) = e 2 (z−1/2) . Note: For a fixed value of x, this is the “mathematician’s version” of the Ztransform described on page 101. Make a change of variable: z = e2πit ⇒ f (x, e2πit ) = =

X

Jn (x)(e2πit )n

n∈Z

X

Jn (x)e2πint ,

n∈Z

Comparing this to the inversion formula in Chapter 1 we get Theorem 5.35. For a fixed x, the n:th Fourier coefficient of the function t 7→ f (x, e2πit ) is equal to Jn (x).

Thus, we can compute Jn (x) by the method descibed in Section 5.3 to compute the coefficients an = Jn (x) (x = fixed, n varies):

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

1) Discretize F (k) = f (x, e 2) Fˆ (m) =

1 N

P∗

2πik N

116

)

− 2πimk N F (k) |k|≤N/2 e

3) Fˆ (m) − Jn (x) =

P

k6=0

am+kN , (Theorem 5.21)

where am+kN = Jm+kN (x).

5.6

One-Sided Sequences

So far we have been talking about periodic sequences (in ΠN ). Instead one often wants to discuss A) Finite sequences A(0), A(1), . . . , A(N − 1) or B) One-sided sequences A(n), n ∈ Z+ = {0, 1, 2 . . .} Note: 5.36. A finite sequence is a special csse of a one-sided sequence: put A(n) = 0 for n ≥ N. Note: 5.37. A one-sided sequence is a special case of a two-sided sequence: put A(n) = 0 for n < 0. Problem: These extended sequences are not periodic. ⇒ We cannot use the Fast Fourier Transform directly.

Notation 5.38. CZ+ = {all complex valued sequences A(n), n ∈ Z+ } Definition 5.39. The convolution of two sequences A, B ∈ CZ+ is (A ∗ B)(m) =

m X k=0

A(m − k)B(k), m ∈ Z+

Note: The summation boundaries are the natural ones if we think that A(k) = B(k) = 0 for k < 0. Notation 5.40. A|n (k) =

(

A(k), 0 ≤ k < n 0,

k ≥ n.

Thus, this restricts the sequence A(k) to the n first terms.

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

117

Lemma 5.41. (A ∗ B)|n = (A|n ∗ B|n )|n Proof. Easy. Notation 5.42. A = 0n means that A(k) = 0 for 0 ≤ k < n − 1, i.e., A|n = 0. Lemma 5.43. If A = 0n and B = 0m , then A ∗ B = 0n+m . Proof. Easy. Computation of A ∗ B 5.44 (One-sided convolution). 1) Choose a number N ≥ 2n (often a power of 2). 2) Define F (k) =

(

A(k), 0 ≤ k < n, 0,

n ≤ k < N,

and extend F to be periodic, period N. 3) Define G(k) =

(

B(k), 0 ≤ k < n, 0,

n ≤ k < N,

periodic extension: G(k + N) = G(k). Then, for all m, 0 ≤ m < n, (F ∗ G)(m) | {z }

=

N −1 X k=0

periodic convolution

=

m X k=0

=

m X k=0

Note: Important that N ≥ 2n.

F (m − k)G(k) F (m − k)G(k) A(m − k)G(k) =

(A ∗ B)(m) {z } |

one-sided convolution

Thus, this way we have computed the n first coefficients of (A ∗ B). Theorem 5.45. The method described below allows us to compute (A ∗ B)|n

(=the first n coefficients of A ∗ B) with a number of FLOP:s which is C · n log2 n, where C is a constant. Method: 1)-3) same as above

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

118

4) Use FFT to compute ˆ N(F \ Fˆ · G(= ∗ G)). 5) Use the inverse FFT to compute F ∗G=

1 ˆ ˆ (F · G)ˇ N

Then (A ∗ B)|n = (F ∗ G)|n . Note: A “naive” computation of A ∗ B|n requires C1 · n2 FLOPs, where C1 is another constant.

Note: Use “naive” method if n small. Use “FFT-inverse FFT” if n is large. Note: The rest of this chapter applies one-sided convolutions to different situations. In all cases the method described in Theorem 5.45 can be used to compute these.

5.7

The Polynomial Interpretation of a Finite Sequence

Problem 5.46. Compute the product of two polynomials: p(x) =

n X

k

ak x

q(x) =

k=0

m X

bl xl .

l=0

Solution: Define ak = 0 for k > 0 and bl = 0 for l > m. Then ! ∞ ! ∞ X X p(x)q(x) = ak xk bl xl =

k=0

|

{z

sums are actually finite

X

ak bl xk+l

k,l

=

X

Pj

l=0

Theorem 5.47.

j

x

j

where cj =

l=0

aj−l bl . This gives

j X l=0

}

(k + l = j, k = j − l)

aj−l bl =

m+n X j=0

cj xj ,

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

119

i) Multiplication of two polynomials corresponds to a convolution of their coefficients: If p(x) = then p(x)q(x) =

Pm+n j=0

n X

k

ak x ,

q(x) =

k=0

m X

bl xl ,

l=0

cj xj , where c = a ∗ b.

ii) Addition of two polynomials corresponds to addition of the coefficients: X p(x) + q(x) = cj xj , where cj = aj + bj .

iii) Multiplication of a polynomial by a complex constant corresponds to multiplication of the coefficients by the same constant. Operation

Polynomial

Addition

p(x) + q(x)

Multiplication by

λp(x)

Coefficients max{m,n}

{ak + bk }k=0

{λak }nk=0

λ∈C

Multiplication p(x)q(x) (a ∗ b)(k) Thus there is a one-to-one correspondence between polynomials ⇐⇒ finite sequences, where the operations correspond as described above. This is used in all symbolic computer computations of polynomials. Note: Two different conventions are in common use: A) first coefficient is a0 (= lowest order), B) first coefficient is an (= highest order).

5.8

Formal Power Series and Analytic Functions

Next we extend “polynomials” so that they may contain infinitely many terms. Definition 5.48. A Formal Power Series (FPS) is a sum of the type ∞ X

A(k)xk

k=0

which need not converge for any x 6= 0. (If it does converge, then it defines an analytic function in the region of convergence.)

120

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

Example 5.49. Example 5.50. Example 5.51.

P∞

xk k=0 k!

P∞

k k=0 x

P∞

converges for all x (and the sum is ex ). converges for |x| < 1 (and the sum is

k k=0 k!x

1 ). 1−x

converges for no x 6= 0.

All of these are formal power series (and the first two are “ordinary” power series). Calculus with FPS 5.52. We borrow the calculus rules from the polynomials: i) We add two FPS:s by adding the coefficients: # # "∞ "∞ ∞ X X X k k [A(k) + B(k)]xk . B(k)x = A(k)x + k=0

k=0

k=0

ii) We multiply a FPS by a constant λ by multiplying each coefficients by λ: λ

∞ X

A(k)xk =

k=0

∞ X

[λA(k)]xk .

k=0

iii) We multiply two FPS:s with each other by taking the convolution of the coefficients: "∞ X k=0

k

A(k)x

#" ∞ X

k

B(k)x

k=0

#

=

∞ X

C(k)xk ,

k=0

where C = A ∗ B.

P k ˜ Notation 5.53. We denote A(x) = ∞ k=0 A(k)x .

Conclusion 5.54. There is a one-to-one correspondence between all Formal Power Series and all one-sided sequences (bounded or not). We denoted these by CZ+ on page 116. Comment 5.55. In the sequence (=”forts¨attningen”) we operate with FPS:s. These power series often converge, and then they define analytic functions, but this fact is not used anywhere in the proofs.

121

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

5.9

Inversion of (Formal) Power Series

˜ Problem 5.56. Given a (formal) power series A(x) = P ˜ verse formal power series B(x) = B(k)xk .

P

A(k)xk , find the in-

˜ Thus, we want to find B(x) so that

˜ B(x) ˜ A(x) = 1, i.e., # #" ∞ "∞ X X B(l)xl = 1 + 0x + 0x2 + . . . A(k)xk l=0

k=0

Notation 5.57. δ0 = {1, 0, 0, . . .} = the sequence whose power series is {1 +

0x + 0x2 + . . .}. This series converges, and the sum is ≡ 1. More generally: δk = {0, 0, 0, . . . , 0, 1, 0, 0, . . .}

= 0 + 0x + 0x2 + . . . + 0xk−1 + 1xk + 0xk+1 + 0xk+2 + . . .

= xk Power Series Sequence xk

δk

˜ B(x) ˜ Solution. We know that A∗B = δ0 , or equivalently, A(x) = 1. Explicitly, ˜ B(x) ˜ A(x) = A(0)B(0) (times x0 )

(5.3)

+[A(0)B(1) + A(1)B(0)]x

(5.4)

+[A(0)B(2) + A(1)B(1) + A(2)B(0)]x2

(5.5)

+...

(5.6)

From this we can solve: i) A(0)B(0) = 1 =⇒ A(0) 6= 0 and B(0) =

1 . A(0)

ii) A(0)B(1) + A(1)B(0) = 0 =⇒ B(1) = − A(1)B(0) (always possible) A(0) 1 [A(1)B(1) + iii) A(0)B(2) + A(1)B(1) + A(2)B(0) = 1 =⇒ B(2) = − A(0)

A(2)B(0)], etc.

we get a theorem: ˜ Theorem 5.58. The FPS A(x) can be inverted if and only if A(0) 6= 0. The

inverse series [A(x)]−1 is obtained recursively by the procedure described above.

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

122

Recursive means: i) Solve B(0) ii) Solve B(1) using B(0) iii) Solve B(2) using B(1) and B(0) iv) Solve B(3) using B(2), B(1) and B(0), etc. This is Hard Work. For example x3 x5 + − ... 3! 5! x2 x4 cos(x) = 1 − + −... 2! 4! 1 sin(x) =??? = sin(x) tan(x) = cos(x) cos(x) {z } | sin(x) = x −

convolution

Hard Work means: Number of FLOPS is a constant times N 2 . Better method: Use FFT. ˜ ˜ Theorem 5.59. Let A(0) 6= 0. and let B(x) be the inverse of A(x). Then, for every k ≥ 1,

B|2k = (B|k ∗ (2δ0 − A ∗ B|k ))|2k

(5.7)

Proof. See Gripenberg. 1 Usage: First compute B|1 = { A(0) , 0, 0, 0, . . .}

Then B|2 = {B(0), B(1), 0, 0, 0, . . .} (use (5.7))

Then B|4 = {B(0), B(1), B(2), B(3), 0, , 0, 0, . . .} Then B|8 = {8 terms 6= 0} etc.

Use the method on page 117 for the convolutions. (Useful only if you need lots

of coefficients).

5.10

Multidimensional FFT

Especially in image processing we also need the discrete Fourier transform in several dimensions. Let d = {1, 2, 3, . . .} be the “space dimension”. Put ΠdN = { sequences x(k1 , k2 , . . . , kd ) which are N-periodic in each variable separately}.

CHAPTER 5. THE DISCRETE FOURIER TRANSFORM

123

Definition 5.60. The d-dimensional Fourier transform is obtained by transforming d successive (=”efter varandra”) “ordinary” Fourier transformations, one for each variable. Lemma 5.61. The d-dimensional Fourier transform is given by xˆ(m1 , m2 , . . . , md ) =

X −2πi(k1 m1 +k2 m2 +...+kd md ) 1 XX N x(k1 , k2 , . . . , kd ). e · · · Nd k k k 1

2

d

Proof. Easy. All 1-dimensional results generalize easy to the d-dimensional case. Notation 5.62. We call k = (k1 , k2 , . . . , kd ) and m = (m1 , m2 , . . . , md ) multiindeces (=pluralis av “multi-index”), and put k · m = k1 m1 + k2 m2 + . . . + kd md (=the “inner product” of k and m). Lemma 5.63. 1 X − 2πim·k N e x(k), Nd k X 2πim·k x(k) = e N xˆ(m). m

xˆ(m) =

Definition 5.64. (F · G)(m) = F (m)G(m) X F (m − k)G(k), (F ∗ G)(m) = k where all the components of m and k run over one period. Theorem 5.65. ˆ (F · G)ˆ = Fˆ ∗ G,

ˆ (F ∗ G)ˆ = N d Fˆ · G. Proof. Follows from Theorem 5.8. In practice: Either use one multi-dimensional, or use d one-dimensional transforms (not much difference, multi-dimensional a little faster).

Chapter 6 The Laplace Transform 6.1

General Remarks

Example 6.1. We send in a “signal” u into an “amplifier”, and get an “output signal” y: u

Black box

y

Under quite general assumptions it can be shown that Z t y(t) = (K ∗ u)(t) = K(t − s)u(s)ds, −∞

i.e., the output is the convolution (=”faltningen”) of u with the “inpulse response” K. Terminology 6.2. “Impulse response” (=pulssvar) since y = K if u = a delta distribution. Causality 6.3. The upper bound in the integral is t, i.e., (K ∗ u)(t) depends only

on past values of u, and not on future values. This is called causality. If, in addition u(t) = 0 for t < 0, then y(t) = 0 for t < 0, and Z t y(t) = K(t − s)u(s)ds, 0

which is a one-sided convolution.

Classification 6.4. Approximately: The Laplace-transform is the Fourier transform applied to one-sided signals (defined on R+ ). In addition there is a change of variable which rotate the complax plane. 124

125

CHAPTER 6. THE LAPLACE TRANSFORM

6.2

The Standard Laplace Transform

Definition 6.5. Suppose that

R∞ 0

e−σt |f (t)|dt < ∞ for some σ ∈ R. Then we

˜ of f by define the Laplace transform f(s) Z ∞ ˜ f (s) = e−st f (t)dt, 0

ℜ(s) ≥ σ.

Lemma 6.6. The integral above converges absolutely for all s ∈ C with ℜ(s) ≥ σ (i.e., f˜(s) is well-defined for such s). Proof. Write s = α + iβ. Then |e−st f (t)| = |e−αt eiβt f (t)| = e−αt |f (t)|

Z

∞

0

≤ e−σt |f (t)|, so Z ∞ −st |e f (t)|dt ≤ e−σt |f (t)|dt < ∞.



0

˜ is analytic in the open half-plane Re(s) > σ, i.e., f˜(s) has Theorem 6.7. f(s) a complex derivative with respect to s. Proof. (Outline) Z ∞ ˜ − f˜(s) f(z) = z−s 0 Z ∞ = Z0 ∞ = 0

As Re(s) > σ we find that 1 2

R∞ 0

e−zt − e−st f (t)dt z−s e−(z−s)t − 1 −st e f (t)dt (put z − s = h) z−s 1 −ht [e − 1] e−st f (t)dt h | {z } →−t as h→0

|te−st f (t)|d < ∞ and a “short”computation (about

page) shows that the Lebesgue dominated convergence theorem can be applied

(show that | h1 (e−ht − 1)| ≤ const. · t · eαt , where α = 21 [σ + ℜ(s)] (this is true R∞ for some small enough h), and then show that 0 teαt |e−st f (t)|dt < ∞). Thus, d ˜ f (s) exists, and ds

d ˜ f (s) = − ds Corollary 6.8.

d ˜ f(s) ds

Z

0

∞

e−st tf (t)dt, ℜ(s) > σ

is the Laplace transform of g(t) = −tf (t), and this

Laplace transform converges (at least) in the half-plane Re(s) > σ.

126

CHAPTER 6. THE LAPLACE TRANSFORM

Theorem 6.9. f˜(s) is bounded in the half-plane ℜ(s) ≥ σ. Proof. (cf. proof of Lemma 6.6) Z ∞ Z ∞ −st |f˜(s)| = | e f (t)dt| ≤ |e−st f (t)|dt 0 Z 0∞ Z ∞ −(ℜs)t = e |f (t)|dt ≤ e−σt |f (t)|dt < ∞. 0

0

Definition 6.10. A bounded analytic function on the half-plane Re(s) > σ is called a H ∞ -function (over this half-plane). R∞ Theorem 6.11. If f is absolutely continuous and 0 e−σt |g(t)|dt < ∞ (i.e., Rt R∞ f (t) = f (0) + 0 g(s)ds, where 0 e−σt |g(t)|dt < ∞), then (f˜′ )(s) = sf˜(s) − f (0), ℜ(s) > σ.

Proof. Integration by parts (a la Lebesgue) gives   −st Z Z T T 1 ∞ −st ′ e −st f (t) 0 + e f (t)dt lim e f (t) dt = lim T →∞ 0 T →∞ −s s 0 | {z } =f˜(s)

1 1 f (0) + f˜′ (s), so s s (f˜′ )(s) = sf˜(s) − f (0).  =

6.3

The Connection with the Fourier Transform

Let Re(s) > σ, and make a change of variable: Z ∞ e−st f (t)dt (t = 2πv; dt = 2πdv) 0 Z ∞ = e−2πsv f (2πv)2πdv (s = α + iω) Z0 ∞ = e−2πiωv e−2παv f (2πv)2πdv (put f (t) = 0 for t < 0) Z0 ∞ = e−2πiωt g(t)dt, −∞

where g(t) = Thus, we got

(

2πe−2παt f (2πt) , t ≥ 0 0

, t < 0.

(6.1)

127

CHAPTER 6. THE LAPLACE TRANSFORM

Theorem 6.12. On the line Re(s) = α (which is a line parallell with the imagi˜ coincides (=sammanfaller med) with the Fourier nary axis −∞ < ω < ∞) f(s)

transform of the function g defined in (6.1).

Thus, modulo a change of variable, the Laplace transform is the Fourier transform of a function vanishing for t < 0. From Theorem 6.12 and the theory about Fourier transforms of functions in L1 (R) and L2 (R) we can derive a number of results. For example: Theorem 6.13. (Compare to Theorem 2.3, page 36) If f ∈ L1 (R+ ) (i.e., ∞), then

lim |f˜(s)| = 0

R∞ 0

|f (t)|dt <

|s|→∞ ℜ(s)≥0

(where s → ∞ in the half plane Re(s) > 0 in an arbitrary manner) Im

Re

Combining Theorem 6.12 with one of the theorems about the inversion of the Fourier integral we get formulas of the type ( Z ∞ e−2παt f (t), t > 0, 1 ˜ + iω)dω = e2πiωt f(α 2π −∞ 0, t < 0. This is often written as a complex line integral: We integrate along the line Re(s) = α, and replace 2πt → t and multiply the formulas by e2παt to get (s = α + iω, ds = idω)

Z α+i∞ 1 f (t) = est f˜(s)ds 2πi α−i∞ Z ∞ 1 = e(α+iω)t f˜(α + iω)idω 2πi ω=−∞

(6.2)

CHAPTER 6. THE LAPLACE TRANSFORM

128

Warning 6.14. This integral seldom converges absolutely. If it does converge absolutely, then (See Theorem 2.3 with the Fourier theorem replaced by the inverse Fourier theorem) the function ( 2πe−2παt f (t), t ≥ 0, g(t) = 0, t<0 must be continuous. In other words: Lemma 6.15. If the integral (6.2) converges absolutely, then f must be continuous and satisfy f (0) = 0. Therefore, the inversion theorems given in Theorem 2.30 and Theorem 2.31 are much more useful. They give (under the assumptions given there) Z α+iT 1 1 ˜ est f(s)ds [f (t+) + f (t−)] = lim T →∞ 2πi α−iT 2

(and we interpret f (t) = 0 for t < 0). By Theorem 6.11, if f is absolutely continuous and f ′ ∈ L1 (R+ ), then (use also Theorem 6.13) 1 f˜(s) = [(f˜′ )(s) + f (0)], s where (f˜′ )(s) → 0 as |s| → ∞, ℜ(s) ≥ 0. Thus, for large values of ω, f˜(α+iω) ≈ f (0) , iω

so the convergence is slow in general. Apart from the space H ∞ (see page

126) (over the half plane) another much used space (especially in Control theory) is H 2 . Theorem 6.16. If f ∈ L2 (R+ ), then the Laplace transform f˜ of f is analytic in

the half-plane ℜ(s) > 0, and it satisfy, in addition Z ∞ ˜ + iω)|2 dω < ∞, sup |f(α α>0

−∞

i.e., there is a constant M so that Z ∞ |f˜(α + iω)|2 dω ≤ M

(for all α > 0).

−∞

Proof. By Theorem 6.12 and the L2 -theory for Fourier integrals (see Section 2.3),

Z

∞

−∞

˜ + iω)|2 dω = |f(α

Z

∞

|2πe−2παt f (2πt)|2 dt (2πt = v) 0 Z ∞ = 2π |e−αv f (v)|2dv 0 Z ∞ ≤ 2π |f (v)|2dv = 2πkf kL2(0,∞) .  0

129

CHAPTER 6. THE LAPLACE TRANSFORM

Converesly: Theorem 6.17. If ϕ is analytic in ℜ(s) > 0, and ϕ satisfies Z ∞ sup |ϕ(α + iω)|2dω < ∞, α>0

(6.3)

−∞

then ϕ is the Laplace transform of a function f ∈ L2 (R+ ). Proof. Not too difficult (but rather long). Definition 6.18. An H 2 -function over the half-plane ℜ(s) > 0 is a function ϕ

which is analytic and satisfies (6.3).

6.4

The Laplace Transform of a Distribution

Let f ∈ S ′ (tempered distribution), and suppose that the support of f is con-

tained in [0, ∞) = R+ (i.e., f vanishes on (−∞, 0)). Then we can define the Laplace transform of f in two ways:

i) Make a change of variables as on page 126 and use the Fourier transform theory. ii) Define f˜(s) as f applied to the “test function” e−st , t > 0. (Warning: this is not a test function!) Both methods lead to the same result, but the second method is actually simpler. If ℜ(s) > 0, then t 7→ e−st behaves like a test function on [0, ∞) but not on (−∞, 0). However, f is supported on [0, ∞), so it does not matter how e−st

behaves for t < 0. More precisely, we take an arbitrary “cut off” function ∞ η ∈ Cpol satisfying

(

η(t) ≡ 1 for t ≥ −1,

η(t) ≡ 0 for t ≤ −2.

Then η(t)e−st = e−st for t ∈ [−1, ∞), and since f is supported on [0, ∞) we can

replace e−st by η(t)e−st to get

Definition 6.19. If f ∈ S ′ vanishes on (−∞, 0), then we define the Laplace transform f˜(s) of f by f˜(s) = hf, η(t)e−st i,

ℜ(s) > 0.

130

CHAPTER 6. THE LAPLACE TRANSFORM

(Compare this to what we did on page 84). Note: In the same way we can define the Laplace transform of a distribution that is not necessarily tempered, but which becomes tempered after multiplication by e−σt for some σ > 0. In this case the Laplace transform will be defined in the half-plane ℜs > σ. Theorem 6.20. If f vanishes on (−∞, 0), then f˜ is analytic on the half-plane ℜs > 0. Proof omitted. Note: f˜ need not be bounded. For example, if f = δ ′ , then g ′ )(s) = hδ ′ , η(t)e−st i = −hδ, η(t)e−st i (δ d −st = e |t=0 = −s. dt (which is unbounded). On the other hand ˜ = hδ, η(t)e−st i = e−st δ(s)

|t=0

= 1.

Theorem 6.21. If f ∈ S ′ vanishes on (−∞, 0), then ) ] i) [tf (t)](s) = −[f˜(s)]′ ℜ(s) > 0 ′ (s) ii) f] = sf˜(s) Proof. Easy (homework?) Warning 6.22. You can apply this distribution transform also to functions, but remember to put f (t) = 0 for t < 0. This automatically leads to a δ-term in the distribution derivative of f : after we define f (t) = 0 for t < 0, the distribution derivative of f is f (0)δ0 | {z }

dervatives of jump at zero

6.5

+

f ′ (t) |{z}

usual derivative

Discrete Time: Z-transform

This is a short continuation of the theory on page 101. In discrete time we also run into one-sided convolutions (as we have seen), and it is possible to compute these by the FFT. From a mathematical point of view the Z-tranform is often simpler than the Fourier transform.

131

CHAPTER 6. THE LAPLACE TRANSFORM

Definition 6.23. The Z-transform of a sequence {f (n)}∞ n=0 is given by f˜(z) =

∞ X

f (n)z −n ,

n=0

for all these z ∈ C for which the series converges absolutely. Lemma 6.24. ˜ i) There is a number ρ ∈ [0, ∞] so that f(z) converges for |z| > ρ and f˜(z) diverges for |z| < ρ.

ii) f˜ is analytic for |z| > ρ. Proof. Course on analytic functions. As we noticed on page 101, the Z-transform can be converted to the discrete time Fourier transform by a simple change of variable.

6.6

Using Laguerra Functions and FFT to Compute Laplace Transforms

We start by recalling some results from the course in special functions: Definition 6.25. The Laguerra polynomials Lm are given by 1 t d
m m −t (t e ), e m! dt

m ≥ 0,

1 t d
m m −t (t e ), e2 m! dt

m ≥ 0.

Lm (t) =

and the Laguerra functions ℓm are given by ℓ, (t) = t

Note that ℓm (t) = e− 2 Lm (t).

Lemma 6.26. The Laguerra polynomials can be computed recusively from the formula (m + 1)Lm+1 (t) + (t − 2m − 1)Lm (t) + mLm−1 (t) = 0, with starting values L−1 ≡ 0 and L1 ≡ 1.

132

CHAPTER 6. THE LAPLACE TRANSFORM

2 + We saw that the sequence {ℓm }∞ m=0 is an ortonormal sequence in L (R ), so that

if we define, for some f ∈ L2 (R+ ), fm =

Z

∞

f (t)ℓm (t)dt,

0

then f (t) =

∞ X

fm ℓm (t) (in the L2 -sense).

(6.4)

m=0

Taking Laplace transforms in this equation we get f˜(s) =

∞ X

fm ℓ˜m (s).

m=0

Lemma 6.27. i) ℓ˜m (s) =

(s−1/2)m , (s+1/2)m+1

P (s−1/2)m ii) f˜(s) = ∞ m=0 fm (s+1/2)m+1 , where fm =

Proof. Course on special functions.

R∞ 0

f (t)ℓm (t)dt.

The same method can be used to compute inverse Laplace transforms, and this gives a possibility to use FFT to compute the coefficients {fm }∞ m=0 if we know f˜(s). The argument goes as follows. ˜ is defined and bounded on Suppose for simplicity that f ∈ L1 (R), so that f(s) C+ = {s ∈ C|Re s > 0}. We want to expand f˜(s) into a series of the type f˜(s) =

∞ X

m=0

fm

(s − 1/2)m . (s + 1/2)m+1

(6.5)

Once we know the cofficients fm we can recover f (t) from formula (6.4). To find the coefficients fm we map the right half-plane C+ into the unit disk D = {z ∈

C : |z| < 1}. We define

s − 1/2 s + 1/2 11+z s= 21−z

z=

1 =1−z s + 1/2 Lemma 6.28.

⇐⇒ and

1 1 sz + z = s − ⇐⇒ 2 2 1 1+z 1 s + s + 1/2 = (1 + )= , so 2 1−z 1−z

133

CHAPTER 6. THE LAPLACE TRANSFORM

i) ℜ(s) > 0 ⇐⇒ |z| < 1 item[ii)]ℜ(s) = 0 ⇐⇒ |z| = 1 iii) s = 1/2 ⇐⇒ z = 0 iv) s = ∞ ⇐⇒ z = 1 v) s = 0 ⇐⇒ z = −1 vi) s = −1/2 ⇐⇒ z = ∞ Proof. Easy. ) is analytic inside the unit disc D, (and bounded Conclusion: The function f˜( 21 1+z 1−z if f˜ is bounded on C+ ). Making the same change of variable as in (6.5) we get ∞ X 1 ˜ 11+z f( )= fm z m . 1−z 21−z m=0

Let us define g(z) = Then

1 ˜ 11+z f( ), 1−z 21−z ∞ X

g(z) =

|z| < 1.

fm z m ,

m=0

so g(z) is the “mathematical” version of the Z-transform of the sequence {fm }∞ m=0

(in the control theory of the Z-transform we replace z m by z −m ). ˜ If we know f(s), then we know g(z), and we can use FFT to compute the coefficients fm : Make a change of variable: Put αN = e2πi/N . Then k g(αN )

=

∞ X

mk fm αN

=

m=0

∞ X

2πimk/N

fm e

m=0

≈

N X

fm e2πimk/N

m=0

(if N is large enough). This is the inverse discrete Fourier transform of a periodic N −1 extension of the sequence {fm }m=0 . Thus, fm ≈ the discrete transformation of N −1 k the sequence {g(αN )}k=0 . We put

k G(k) = g(αN )=

k 1 ˜( 1 1 + αN ), f k k 2 1 − αN 1 − αN

ˆ and get fm ≈ G(m), which can be computed with the FFT.

CHAPTER 6. THE LAPLACE TRANSFORM

134

Error estimate: We know that fm = gˆ(m) (see page 115) and that gˆ(m) = 0 for m < 0. By the error estimate on page 108 we get ˆ |G(m) − fm | = (where we put fm = 0 for m < 0).

X k6=0

|fm+kN |

Bibliography [Gri04] Gustaf Gripernberg, Fourier-muunnosten perusteet (lecture notes), http://math.tkk.fi/teaching/fourier/, 2004. [GW99] C. Gasquet and P. Witomski, Fourier analysis and applications, Texts in Applied Mathematics, vol. 30, Springer-Verlag, New York, 1999, Filtering, numerical computation, wavelets, Translated from the French and with a preface by R. Ryan. MR MR1657104 (99h:42003) [Wal02] David F. Walnut, An introduction to wavelet analysis, Applied and Numerical Harmonic Analysis, Birkh¨auser Boston Inc., Boston, MA, 2002. MR MR1854350 (2002f:42039)

135